Generative AI Chat Interface — Visual Designer AI Tab UX

Status: Research (revised direction) | ClickUp: 868hu2bq6 Sibling idea: milo-multi-agent-interface — kept intact, separate direction Direction update (2026-07-09): UWW funding scope + CEO re-rulings — see 2026-07-09 Direction Update. The conversational guide is now the primary homeowner flow.

Overview

Transform the Visual Designer's AI tab from a near-empty chat-prompt panel into a guided, step-by-step design workflow that mirrors how professional landscape designers and architects actually work. The default state needs to be immediately useful with zero typing. Replace the open prompt as the primary action with a structured, visual, click-driven flow informed by a landscape "design system" chosen up front.

The AI tab is one panel inside the Visual Designer, not a full surface takeover. It must do everything described in this brief while living entirely within that panel and coexisting with the other tabs (Plants, Objects, Images, Draw, Clone) that share the canvas. See Surface Constraints immediately below — that section governs everything else in this brief.

Surface Constraints

This is the most important section of the brief. Earlier iterations of this idea's UX assumed the AI tab owned breadcrumb, sub-tabs, and a left-rail stepper outside the panel. That is wrong. All AI tab UX must be designed inside the panel.

The AI tab is a panel, not a page. The current Visual Designer chrome looks like this (left to right, top to bottom):

• Top chrome (owned by Visual Designer, not the AI tab): home icon, File menu, undo/redo, sync icon, property name (e.g., "Brimhall's"), hamburger menu, Share button. The AI tab does not add a breadcrumb, sub-tab strip, or any other element here.
• Left tab strip (owned by Visual Designer): vertical icon strip with AI, Plants, Objects, Images, Draw, Clone. AI is the active tab in this brief. The strip is a peer-tab navigator, not part of the AI panel.
• AI panel (owned by this brief): docked to the right of the tab strip and to the left of the canvas. Approximately 290–320px wide on desktop; on mobile, approximately the same working width (panel-style sheet). Internal layout:
  • Panel title bar ("AI Edit" with a close X) — internal, not the Visual Designer top bar.
  • Generation history (vertical, mid-panel): each entry shows a short label plus before/after thumbnails (e.g., "Add a stone walkway", "change to nighttime"). This is the existing version-history UI, elevated.
  • Prompt input (bottom of panel): existing free-text input, paid-tier-gated, retained.
  • New AI tab UX added by this brief — design-system pick, questionnaire, the five-step guided flow, drag-and-drop staging handles, references, etc. — all live inside this panel at the same approximate dimensions.
• Canvas (shared real estate): the canvas is shared across all tabs. Any overlays the AI tab places on the canvas (object outlines for object-deletion, ghost staging for drag-and-drop, plant labels for the Views step) must be designed to coexist with overlays from Plants, Draw, and Clone. The AI tab does not own the canvas.

Implications for downstream UX work:

• No top-bar breadcrumbs ("Property > Area"), no top-bar sub-tabs ("Photos / Designs / AI"), no left-rail stepper outside the panel. Earlier wireframes that introduced any of these are wrong.
• The "system pill" (e.g., "Cottage ▾"), the step indicator, the Apply/Cancel/Skip controls, the questionnaire, the prompt mode toggle, and references all live within ~290–320px of horizontal space.
• Mobile is the same working space — design once for the panel, not twice for desktop and mobile.
• The Property Design Profile (see Data Model Notes) may be edited via a panel-internal modal/sheet that appears inside or over the panel — never via a full-page route.

Problem

The current AI tab forces users to type prompts to do anything. Three problems with that:

1. Most users don't know what to type. Empty prompt + a couple of suggestion chips doesn't unlock the imagination — it triggers blank-page paralysis.
2. Typing is slow. Even users who know what they want would prefer to click than to write a paragraph.
3. Free-text prompts don't build trust. Users can't tell whether the AI will do what they meant. They want preview, comparison, and incremental steps — not a black-box single shot.

Competitor "style template" tools (Neighborbrite, Garden AI) reduce friction with one-click style transformations, and Neighborbrite's paid tier adds a "Create your style" prompt-and-element editor. But these are one-and-done collapses of the entire design into a single AI generation. Real landscape design happens in stages.

Strategic Fit

This is the default surface for AI-assisted design in SimplyScapes. It must:

• Lower the cost-of-first-action (no prompt-writing required to make progress).
• Build trust by mirroring a recognizable professional workflow.
• Coexist with — and feed into — the more capable MILO multi-agent interface (sibling idea, kept intact).
• Support pro users who want to type prompts directly while keeping that path optional and paid-tier-gated.

2026-07-09: This surface's strategic weight increased — the conversational, question-driven guide was promoted to the primary homeowner flow by CEO re-ruling. See the 2026-07-09 Direction Update.

2026-07-09 Direction Update: UWW Funding Scope and CEO Re-Rulings

Recorded by Dan from the 2026-07-06 Utah Water Ways (UWW) funding meeting (Scott Brady + Cynthia Bee, UWW) and CEO rulings from the 2026-07-09 session. This section records product scope and direction only — commercial terms of the UWW engagement are deliberately not documented in this public repo.

CEO re-ruling: conversational guide promoted to primary homeowner flow

CEO re-ruling (2026-07-09). The conversational, question-driven guided design flow developed in this idea is promoted to the primary homeowner flow. This supersedes the scenario-rebate-compare recommendation, which ranked the H1 Linear Wizard as the spine and scored the H2 Conversational Guide as an "Alternate front-end" (A/B alternate). That H1-primary ranking is superseded for the homeowner flow as of this ruling; the conversational paradigm leads. (No document inside this idea ranked the conversational flow as an alternate, so there is nothing to strike locally — the superseded ranking lives in the rebate-compare exploration and is corrected here by dated supersession.)

Cynthia Bee's framing from the meeting captures the paradigm: "AI is prompting them versus them prompting AI… we're the guide." The system asks; the homeowner answers and chooses. This is the same zero-typing, click-driven principle this brief already established — now elevated from "default AI-tab surface" to the primary way a homeowner designs.

UWW extension scope

Four extensions to the guided walkthrough, proposed as UWW-funded scope on top of the flow this brief defines:

1. Design-intent interview questions. The current sample questionnaire covers usage, maintenance, water, and vibe — it lacks design-intent questions: "Do you want a back patio? How many people do you need to seat? What's your activity level?" These are asked step by step as decisions inside the walkthrough, not batch-front-loaded — each answer immediately shapes the next design decision.
2. "Before you begin" prep checklist stage. A new stage shown before the walkthrough starts: test your water pressure, look up easements, gather property information. Gets the homeowner physically and informationally ready before any design decision is asked of them.
3. Video-embed slots at key decision points. The walkthrough gets embed capacity for short instructional videos at decision points — capacity first, content later ("we know we won't have all the videos yet"). Cynthia's Teachable class inventory (Intro to LocalScapes, LocalScapes University, Design Workshop, Irrigation Workshop, Flip Your Strip) is researched in the LocalScapes five-step-process research report — Classes and Education and is the seed content pool.
4. $/$$$ cost-consciousness prompts. Walkthrough steps that surface cost posture ("here's the most cost-effective path") and hand off to the estimate surface for the actual numbers. These prompts are guidance inside the flow, not a pricing engine of their own.

CEO re-ruling: References / inspiration photos un-deferred

CEO re-ruling (2026-07-09). ~~References confirmed deferred by the CEO on 2026-06-04 — out of the shipped templates phase~~ (deferral recorded in the AI Design Templates brief — Open questions for CEO): superseded 2026-07-09 — the References / inspiration-photo subsystem is un-deferred as UWW-funded scope. The References design in this brief (Scope, Data Model Notes) carries forward as specced. The un-deferral applies to References specifically; the other items in the 2026-06-04 deferral list (Views, Property Design Profile, Site Analysis, 5-step flow) are not re-ruled by this note.

CEO framing: "They can just look at it and go, I like the way that looks… turn something they just saw into a plan for their own yard."

Open scoping question — "apply that look" style transfer vs. Jobs CEO Decision 5. One-click apply-that-look style transfer (turning an inspiration photo into a generated design for the user's own yard in one action) remains net-new model work, and it is positioned against — not silently overriding — CEO Decision 5 in the Jobs brief, which ruled one-click image→design to be a photo-seeded blank canvas, explicitly NOT an AI auto-render. Whether "apply that look" is (a) a distinct, later generation feature that coexists with the photo-seeded canvas ruling, or (b) grounds to revisit Decision 5, is an open question for the CEO — do not build or spec style transfer until it is explicitly re-ruled.

Candidate architecture: context-fed guide (CEO, 2026-07-09)

Candidate — not ratified. CEO framing: "guide what's possible — don't close down discovery." The architecture sketch below records the direction of thinking so spec work can evaluate it; it is not a committed design.

• Milo as the conversational surface. The in-app Milo AI widget (web PR #250, open at time of writing) is the candidate host for the conversational guide, rather than a separate chat implementation inside the AI panel.
• Instance-specific wiki in a Jobs-tab side panel. Each whitelabel instance gets a wiki surface (side panel on the Jobs tab) whose content is ingested into Milo — instance operators author the knowledge that steers their homeowners' guided experience. 2026-07-09 (later same day): scoped as its own product-level idea — Wiki Authoring Platform (partner authoring interface, Jobs-tab reader side panel, public wiki pages, Milo corpus ingestion) — agreed as a new UWW funding ask. That idea is the content backbone of the UWW extension scope above.
• Markdown (MD) context fields — internal and user-facing — on users, clients, properties, and jobs, feeding Milo's brain and steering the linear experience. This aligns with the platform SSFM decision (GFM markdown as agent-native source of truth) — see Agent-Native Rich Text Format (SSFM) research.
• Geo-overlapping entities contribute jurisdiction context. State / District / Municipality / HOA entities with geographic boundaries (PostGIS) overlap a property's location and contribute their context (rules, programs, plant guidance) into the same context pipeline that feeds the guide.

Scope

In scope (this idea):

• The AI tab panel inside the Visual Designer, applied to edits on an existing rendered image only (not from-scratch design). All UX lives inside the panel per Surface Constraints.
• Auto-classification at site analysis time — extends the existing EXIF + AI vision pipeline (yard type, orientation, climate zone, size, removable objects) with two new classifications:
  • Home architectural style (Craftsman, Ranch, Modern, Mediterranean, Tudor, Colonial, Mid-Century, Farmhouse, Spanish, Contemporary, etc.). Used to soft-recommend matching design systems on the picker — top 2–3 cards get a small "Recommended for your home" ribbon — and to inform questionnaire defaults.
  • Existing landscape style (Cottage-leaning, Naturalistic-leaning, Formal, Bare/Lawn-only, Overgrown/Untended, Xeriscape-leaning, etc.). Used to inform the "what to keep" decision in Step 1 (Prepare site).
  • Both classifications are confirmable / correctable by the user. Auto-classification is a starting point, never a verdict.
• Site-aware lead-in: EXIF + AI vision + the two new classifications feed into a short click-through questionnaire (with optional text answers).
• Pre-baked landscape design systems users pick from up front (e.g., Cottage, Modern, Naturalistic, Formal Traditional, Xeriscape, Mediterranean, Tropical, Coastal, Japanese Zen, Modern Farmhouse, Prairie/Native, Rustic — twelve recommended in research).
• Multi-step guided workflow with skippable, non-linear steps:
  1. Prepare the site — remove unwanted items (overgrown shrubs, old trees, vehicles, clear a flower bed, clear all plants). Existing-landscape-style classification informs default keep/remove suggestions.
  2. Define spaces and layout — reduce grass to a central shape, add deep planter beds, hardscapes, structures.
  3. Add base plantings — trees, foundation shrubs, perennials, bedding plants, groundcovers. Drag-and-drop from the existing plant and object library; users continue placing until they hit "Apply", at which point the staged elements flatten into the generated image.
  4. Decor + finishing touches — lighting, accessories, structures.
  5. Views — plant labels overlay, camera angle changes (top view, side view, etc.).
• References inside the panel — a "References" section in the AI panel where the user can pull in (a) alternate base views of the same area (front view, side view, top-down survey), (b) other visual designs from elsewhere on the same property (front yard hardscape feeding context to a backyard design), and (c) aspirational inspiration images uploaded by the user. References are optional context the prompt-construction layer may include in any generation. See Data Model Notes for entity relationships. (2026-07-09: un-deferred as UWW-funded scope, superseding the 2026-06-04 templates-phase deferral — see the Direction Update.)
• Optional text-prompt entry available at any step (paid-tier gated).
• Skippable steps and free reordering.
• Tight integration with the existing version-comparison/revert UI (already in the panel — generation history with before/after thumbnails). Reverting to a previous render is part of the design loop, not a footnote.

Out of scope (preserve as separate ideas):

• The MILO multi-agent cascading-bubble interface (sibling idea — different bet on UX).
• From-scratch design / property onboarding (handled elsewhere).
• 3D rendering, Site Planner, hardscape design as standalone capabilities.
• Editing the Property Design Profile from outside the Visual Designer (e.g., a separate property management page) — that's a future capability, not v1.

Key Constraints

• Panel-only surface (CRITICAL). All AI tab UX lives inside the ~290–320px panel docked between the Visual Designer's tab strip and the canvas. Mobile is approximately the same working space. The Visual Designer's top chrome (home, File menu, undo/redo, sync, property name, hamburger, Share) and tab strip (AI, Plants, Objects, Images, Draw, Clone) are not editable from this brief. The canvas is shared real estate across tabs. See Surface Constraints for the full statement.
• Default state must be immediately useful with zero typing. A new user opening the AI tab should see a clear next action — not a blinking text field.
• Workflow must be non-linear. Users can skip, jump back, swap steps, or branch into parallel concepts via the version history.
• Free-text prompt entry is a paid-tier feature, not the primary path. Click-driven UI is the default.
• Existing version-history / revert UI is preserved and elevated into a first-class part of the workflow (used to spin up multiple concept variants in a non-linear-but-linear way). It already lives mid-panel; this brief keeps it there.
• Drag-and-drop leverages the existing plant and object library — not a new asset system. Staged objects flatten on "Apply".
• Site analysis auto-classifies, never dictates. Yard type, orientation, climate, size, removable objects, home architectural style, and existing landscape style are all auto-detected at site analysis time and surfaced as confirmable/correctable starting points — never as locked-in verdicts.
• Property Design Profile is the source of truth. A property's design system, questionnaire answers, preferences, and home-style classification live as a standalone Property Design Profile entity, referenced by N visual designs (front yard, backyard, parkstrip, etc.). See Data Model Notes.

Hypothesis

A guided, click-driven workflow rooted in a chosen landscape design system, delivered inside the existing AI tab panel and powered by a property-level Profile that propagates across all of a property's visual designs, will let users produce more refined, multi-step designs than competitor one-shot style transforms — without learning to write prompts and without reconfiguring the system for every area of the same property. Pro users who want prompt control still have it; novices have a clear visual path.

Implementation Notes (for spec phase)

• Panel real estate is the binding constraint. Every component added by this brief must fit within ~290–320px of horizontal panel width. Patterns to consider in spec: vertical stacking, accordion sections, modal/sheet panels that overlay the panel itself (not the canvas) for short-lived editors (e.g., the questionnaire, the Property Design Profile editor, the References manager).
• The multi-step image-generation pipeline is a strong fit for a durable workflow runtime (e.g., Vercel Workflow DevKit). Each step is a long-running generation; users may pause and resume across sessions; failures need retries; users may cancel mid-step. Capture this in the spec, not now.
• Site analysis is now a one-time-per-property call (not one-time-per-session) because results feed the Property Design Profile. Outputs include the existing classifications (yard type, orientation, climate zone, size, removable objects) plus the two new classifications (home architectural style, existing landscape style). All are persisted on the profile and confirmable by the user.
• Auto-classification confidence must be returned with each label. The spec should require the AI vision call to return per-label confidence so the UI can decide when to surface "Looks like Craftsman — confirm?" vs. silently apply.
• Drag-and-drop "flatten on apply" implies a staging layer above the generated image that compiles into a single prompt + base image when the user commits. Staging overlays render on the shared canvas and must coexist with overlays from other tabs. The spec will need to nail down the staging-to-prompt contract and the cross-tab overlay coordination.
• References add a new input class to prompt construction. When a generation runs, the prompt-construction layer optionally includes reference images: alternate base views of the same area, other visual designs from the same property, or aspirational images from the user's library. The spec must define how many references are acceptable per generation (cost/latency tradeoff), how reference intent is described to the model, and how user selection of references is persisted on the resulting generation history entry.

Data Model Notes (for spec phase)

These notes sketch the entity relationships the spec must formalize. They are product-level (what entities exist, how they relate), not implementation-level (table names, FK column names, GraphQL shape — those are the spec's job).

Entities

Property

• A real-world address.
• Has one Property Design Profile (1:1, may be lazily created).
• Has many Visual Designs (each scoped to an area of the property).
• Has a property-level Reference Library (uploaded inspiration images, plus references promoted from individual visual designs).

Property Design Profile (new entity introduced by this brief)

• Standalone, property-scoped.
• Holds the chosen design system (e.g., "Cottage"), the questionnaire answers, user preferences (e.g., maintenance load, water use, color preferences), the home architectural style classification, and the climate context (zone, orientation, yard-type defaults at the property level).
• Editable in one place; changes propagate to every Visual Design that inherits from it.
• Surfaced as a panel-internal modal/sheet inside the AI tab — not a separate page.

Visual Design

• Scoped to one area of a property (front yard, backyard, parkstrip, side yard, garden, patio, etc.).
• Inherits from the Property Design Profile by default. Each attribute on the visual design is either inherited or overridden.
• Per-area customizations are diffs against the profile, not duplications. Example: a backyard with design_system: "Tropical" overriding the property's design_system: "Cottage" is one diff, not a full clone of the profile.
• Has many Base Images (the alternate base views of the same area: front view, side view, top-down survey).
• Has many Generations (the existing version-history entries).
• Has many Active References per generation (the references selected at generation time).

Reference Image

• Three subtypes:
  1. Alternate base view — additional photographs of the same area at different angles.
  2. Property reference — image attached at the property level (e.g., the front yard's hardscape used as context when designing the backyard) and selectable from any visual design on that property.
  3. Aspirational reference — uploaded inspiration photos, attached to the property profile and/or directly to a visual design.
• May be tagged for intent (e.g., "color palette", "plant placement", "hardscape style"). Tagging is a v2 question (see Research Question 8).

Inheritance & override behavior

• A new visual design defaults to inheriting the full Property Design Profile. The questionnaire is not re-asked unless the user explicitly asks to customize for that area.
• The "system pill" (or its panel-only equivalent) on a visual design switches the area's system. The UI must show whether the active system is inherited or overridden.
• Editing the Property Design Profile updates every Visual Design that inherits from it. Visual Designs that override an attribute keep their override.
• The user can "propagate up" (push a change made on a visual design back to the property profile) — UX for this is Research Question 7.

Why this matters for the spec

• The Profile entity changes the data shape: today's design lives in one place; tomorrow's lives in two — the profile and the per-design diff. The spec must define the entity, the inheritance resolution rule, and the migration path for existing designs.
• The References subsystem changes the prompt-construction layer. Generations now take an N-tuple of input images (base + N references) plus the system fragment plus the step prompt. The spec must define the per-generation reference budget (cost/latency).
• The two new auto-classifications add fields to the site-analysis output and to the Profile. The spec must define the schema and the confidence-threshold defaults.

Research Questions (this round)

Status note (2026-05-01). Questions 1–4 below were answered by the prior research round — see report.md Part V. Questions 5–8 are new, raised by the revised direction (panel-only surface, expanded auto-classification, Property Design Profile, multi-image references). The next research/UX round should focus on 5–8.

1. ~~Professional designer workflow.~~ Answered in report.md § V.A.
2. ~~Adjacent AI design tools with onboarding flows.~~ Answered in report.md § V.B.
3. ~~The "design system" concept applied to landscaping.~~ Answered in report.md § V.C.
4. ~~Sample artifacts.~~ Delivered in ux/sample-questionnaire.md and ux/design-systems.md.
5. Panel-constrained interaction patterns. What proven patterns exist for fitting a multi-step guided workflow, a system pill, a step indicator, drag-and-drop staging, generation history, a references manager, and a prompt input into a 290–320px panel? Look at IDEs (VS Code side panels), design tools (Figma right panel, Photoshop right panel), and AI assistants (Cursor side chat). Identify density patterns, accordion vs. tabs vs. sheets, and what to push to overlays vs. keep inline.
6. Auto-classification UX patterns for home architectural style and existing landscape style. How should the panel surface confidence-scored classifications? When is a passive label OK and when does the user need an explicit confirm step? What's the correction UX inside ~300px (chip menu, dropdown, modal sheet)? Are there comparable patterns in interior-design tools (Modsy, Havenly) for room-style auto-detection? What confidence threshold should trigger a soft "Recommended for your home" ribbon on picker cards vs. a hard recommendation?
7. Property Design Profile inheritance UX. When a user is in the Backyard's AI tab and wants to override the property profile (e.g., "Backyard uses Tropical, overriding the property's Cottage"), how is the override surfaced? Inheritance indicator (e.g., a small "inherits from property profile" link), break-link affordance, and "propagate this change to the property" reverse path. Look at Figma component overrides and CSS cascade UIs for prior art.
8. Multi-image reference patterns inside the panel. How should references be added (drag-in from canvas, picker from property library, upload), grouped (alternate views vs. property references vs. aspirational), weighted (does the user say "use this for color palette" vs. "use this for plant placement"?), and surfaced per-generation (which references were active for this output)? Reference: how Midjourney/Stability handle reference images and how Adobe Firefly handles "Style Reference" panels.

Output Artifacts (this round)

Note (2026-05-01). The prior round delivered research Part V and the full UX set (flow-diagram.md, wireframes.md, design-systems.md, sample-questionnaire.md, claude-design-prompt.md). Those deliverables assumed an out-of-panel surface (top-bar breadcrumb, sub-tab strip, left-rail stepper) that has now been ruled out. This next round must redo the UX set against the panel-only constraint and add research and UX coverage for the three other revisions (auto-classification, Property Design Profile, multi-image references). The previously-shipped UX artifacts are not deleted — they remain as historical reference for the workflow shape and design-system content, both of which carry forward.

Research:

• Append research/report.md with a new Part VI covering Research Questions 5–8: panel-constrained interaction patterns, auto-classification UX, Property Design Profile inheritance UX, and multi-image reference patterns.
• research/supporting/panel-constrained-design-tools.md — IDE/design- tool side-panel patterns at 280–340px width (VS Code, Figma, Photoshop, Cursor, Adobe Firefly).
• research/supporting/auto-classification-ux.md — interior-design and architecture analogues for room-style and home-style auto- detection; confidence-threshold patterns.
• research/supporting/profile-inheritance-ux.md — Figma component overrides, CSS cascade UIs, and other inheritance/override prior art.

UX (revised against panel-only constraint):

• ux/wireframes-v2.md — full wireframe set redone inside the panel. Replaces — does not append to — the v1 wireframes for the path forward. Must cover: design-system picker (panel-grid), site analysis result + auto-classification confirms, the five-step flow, drag-and-drop staging with shared-canvas overlay, references manager, Property Design Profile editor (panel-internal modal/sheet), generation history, paid prompt mode.
• ux/flow-diagram-v2.md — updated flow diagram reflecting the Property Design Profile entry/editing paths and the references workflow.
• ux/auto-classification-spec.md — spec for the home-architectural- style and existing-landscape-style classifiers from a UX perspective: label set, confidence thresholds for soft vs. hard recommendations, correction patterns.
• ux/profile-and-references.md — UX for the Property Design Profile modal and the References manager inside the panel. Includes inheritance indicators, override break-link affordances, and the reference picker (alternate views / property references / aspirational).
• ux/claude-design-prompt-v2.md — refreshed paste-ready Claude Design brief that bakes in the panel-only constraint, the new classifications, the Profile entity, and references. Headline deliverable for this round.

Version History

| Date & Time (MT) | Author | Summary | |--------------------------|---------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | 2026-03-09 — MT | Dan | Initial brief stub. | | 2026-05-01 02:00 PM MT | Jarvis | Pivoted direction: chat interface → guided step-by-step workflow with landscape design systems. Captured CEO direction; scoped to image edits; preserved MILO sibling idea; flagged durable workflow as spec consideration. | | 2026-05-01 07:30 PM MT | UX Designer | Landed UX deliverables: ux/flow-diagram.md (mermaid + 8-question resolutions), ux/wireframes.md (10 screens with image prompts), ux/design-systems.md (12 systems + starter prompt fragments), ux/sample-questionnaire.md (7+1 questions + modifiers), ux/claude-design-prompt.md (paste-ready Claude Design brief, ~6k words). Headline deliverable is the Claude Design prompt — Scott will paste it into claude.ai/design to generate an interactive prototype. | | 2026-05-01 09:30 PM MT | Dan | Four critical revisions captured. (1) Surface Constraints — added a top-billed section establishing the AI tab as a single ~290–320px panel docked between the tab strip and the canvas, with all UX inside the panel; flagged earlier wireframes' breadcrumb / sub-tabs / left-rail stepper as wrong. (2) Auto-classification expanded — added home architectural style and existing landscape style to the EXIF + AI vision pipeline; confirmable/correctable, used to soft-recommend matching design systems and inform the questionnaire. (3) Property Design Profile — introduced as a standalone, property-scoped entity that holds the chosen design system, questionnaire answers, preferences, and home-style classification; visual designs inherit by default with per-area diffs as overrides; documented in new "Data Model Notes (for spec phase)" section. (4) Multi-image references — added a References section to scope (alternate base views, property references, aspirational uploads) and to the data model. Refreshed Output Artifacts to redo the UX set against the panel constraint (wireframes-v2, flow-diagram-v2, profile-and-references, auto-classification-spec, claude-design-prompt-v2) and to add a Part VI research round (Q5–8). Strikethroughed answered Research Questions 1–4 with deep links to the existing report; added Q5–8. | | 2026-05-02 12:20 AM MT | UX Designer | Round 2 UX deliverables landed against the panel-only constraint. Revised: ux/wireframes.md (all 10 prior screens redrawn as panel-only ~290–320px wide; 5 new screens added — panel-with-profile, profile editor sheet, References section variants, system pill inheritance popover, system picker with auto-classification ribbons; Step 1 now uses auto-keep-assessment from existing-landscape classification; site analysis surfaces home-style + landscape-style with confidence-tier UI; cross-tab canvas overlay z-order rules captured in Engineering Reference). ux/flow-diagram.md (happy-path mermaid now includes profile load/skip branch, migration prompt, auto-classification confidence-tier branches, profile editor sheet, system pill override popover, references add-sheet, mid-panel generation history; resolved all 12 round-2 open questions from research § Part VI Section E). ux/design-systems.md (added Profile composition subsection — system resolution as visualDesign.systemOverride ?? property.profile.system, references-as-summaries pattern, profile questionnaire-answers as soft modifiers; updated note 5 for panel-internal system pill; added note 6 about profile-level rewrite cascading). ux/sample-questionnaire.md (compressed for 320px panel — chip-pick primary, image-pick uses 1-column horizontal-thumbnail rows; Q5 home vibe auto-skipped at high confidence and pre-filled at medium; site summary template adds home-style + existing-landscape-style as bolded auto-classified tokens; both editable via chip menu). ux/claude-design-prompt.md (heavy rewrite — Surface Constraints section first, all 10 prior screens rewritten as panel-only, 5 new screens added covering profile editor sheet/references add-sheet/inheritance popover/system picker, mobile bottom-sheet variant; ~9100 words). New file: ux/property-design-profile.md — UX specification for the Profile entity covering five panel surfaces (header strip, system pill, editor sheet, system picker integration, propagate-up reverse path), inheritance resolution UI rules, migration UX for properties with prior designs, refresh-on-open flow for inheriting designs, profile lifecycle states, open questions for spec. Updated _meta.yaml ux list and updated: 2026-05-02. | | 2026-07-09 06:25 PM MT | Dan | 2026-07-09 Direction Update recorded from the 2026-07-06 UWW funding meeting (Scott Brady + Cynthia Bee) and 2026-07-09 CEO rulings (public-safe: no commercial terms). (1) CEO re-ruling — conversational, question-driven guide promoted to PRIMARY homeowner flow, superseding the scenario-rebate-compare H1-wizard-primary ranking (H2 was scored "Alternate front-end"); Cynthia's framing quoted ("AI is prompting them versus them prompting AI… we're the guide"). (2) UWW extension scope — design-intent interview questions (patio? seating count? activity level?) decided step by step; "Before you begin" prep checklist stage (water pressure, easements, property info); video-embed slots at decision points (capacity first, content later; Teachable inventory linked); $/$$/$$$ cost-consciousness prompts handing off to the estimate surface. (3) CEO re-ruling — References/inspiration photos UN-DEFERRED as UWW-funded scope (strikethrough supersession of the 2026-06-04 templates-phase deferral, linked); "apply that look" style transfer recorded as net-new model work positioned AGAINST (not overriding) Jobs CEO Decision 5 (photo-seeded blank canvas) — open CEO question. (4) Candidate architecture (not ratified, "guide what's possible — don't close down discovery"): Milo widget as conversational surface (web PR #250), instance wiki in Jobs-tab side panel ingested into Milo, MD context fields on users/clients/properties/jobs (aligned with SSFM, linked), geo-overlapping State/District/Municipality/HOA entities via PostGIS. Added dated pointers in header, Strategic Fit, and the Scope References bullet. Later same day: cross-linked the new wiki-authoring-platform idea (new UWW funding ask) as the wiki bullet's dedicated idea and the content backbone of the UWW extensions. |

Generative AI Chat Interface: API Design, Interaction Model & Pricing

ID: SS-RR-2026-003 | Date: 2026-03-09 | Status: draft ClickUp: Generative AI Chat Interface Plan: SS-RP-2026-003 Domain: Generative AI / Visual Designer

TL;DR

SimplyScapes' v1 AI inpainting works but is single-turn: prompt in, image out. The v2 chat interface transforms this into a conversational design partner that can ask follow-up questions, present choices from the 2,500+ object library with thumbnails and pricing, accept user markup/annotations for spatial guidance, and answer landscaping knowledge questions — all through a two-phase intent routing architecture that uses cheap text classification (Gemini 2.5 Flash, ~$0.001/request) before expensive image generation ($0.039/image). No competitor in the landscape vertical offers this combination. The patent landscape is favorable (overall Low-Moderate risk, no landscape-specific AI patents exist), and the $995 gap between free AI inspiration tools and human design services represents a massive underserved market. A credit-based pricing model (10 credits = 1 image generation, free tier 50 credits/mo) aligns with the 126% YoY growth in credit-based SaaS pricing. The full architecture — intent routing, credit system, conversation persistence, object library integration, and markup-guided generation — is buildable in three 4-6 week phases using Gemini, Hasura, and the existing Next.js stack.

---

Part I: The Idea

1. What We're Exploring

SimplyScapes has a working v1 of AI-powered image generation in its Visual Designer — users type a prompt, Gemini generates or modifies the design background, and the result is displayed with edit history and revert. The interaction is single-turn: prompt in, image out.

The v2 evolution introduces a conversational AI layer. Instead of always generating immediately, the AI can ask follow-up questions ("Which style of fountain?"), present choices from the SimplyScapes object library (2,500+ plants, hardscape objects, materials), accept user markup/annotations to guide placement, and answer knowledge questions — all within a chat interface that already exists in the designer UI.

The core challenge is threefold: (1) Intent routing — how does the AI decide whether to generate, clarify, or answer? (2) Pricing — how do you meter and monetize a conversational AI experience with fine-grained credits? (3) API design — what does a general-purpose endpoint look like that handles all these interaction types while staying compatible with Gemini's API?

2. Why It Matters

This is SimplyScapes' transition from "AI as a feature" to "AI as the interaction layer." The current mode dropdown (Edit, Plant Selection, Ask, Finalize, Erase) already signals this ambition — the UI is ahead of the backend.

The business case is direct: AI-assisted design is the primary differentiator against desktop-bound competitors (ProLandscape, DynaSCAPE, Idea Spectrum) and the key retention driver for professionals who need to produce designs quickly in the field. Getting the interaction model right means landscapers spend less time crafting prompts and more time reviewing options — which is exactly the workflow that converts free users to paid subscribers.

The credit system is the monetization mechanism. Done well, it creates predictable revenue from AI usage while keeping the free tier generous enough to demonstrate value. Done poorly, it creates friction that drives users to competitors or to using Gemini directly.

The API design has long-term implications beyond the chat interface. A well-abstracted endpoint becomes the foundation for AI-powered proposals, plant recommendations, maintenance schedules, and any future AI capability — all through the same contract.

---

Part II: Research Findings

3. Intent Routing & Function Calling

The core architectural question: how does the system decide whether a user message should trigger image generation, a follow-up question, or a text answer?

3.1 Gemini's Function Calling Mechanism

Gemini's function calling API provides the foundation for intent routing. You declare tools (functions) with JSON schemas describing their names, parameters, and descriptions. When a user message arrives, Gemini analyzes it against the declared tools and either calls a function or responds with natural language.

Four calling modes control behavior:

| Mode | Behavior | Use Case | |------|----------|----------| | AUTO (default) | Model decides between function call or natural language | General-purpose routing — recommended for SimplyScapes | | ANY | Model must call a function from declared set | Forced classification — useful when you always want structured output | | NONE | Function calls disabled | Ask mode when you only want text answers | | VALIDATED (preview) | Ensures either valid function call or natural language | Stricter validation for production |

For SimplyScapes, the recommended approach is a two-phase architecture:

1. Phase 1 — Intent Classification (function calling, text-only model). Send the user's text prompt + mode context to a cheap text model (Gemini 2.5 Flash or 3 Flash) with tool declarations. The model decides which tool to call: generate_image, ask_followup, search_objects, search_plants, answer_question, or request_markup. This call is inexpensive — text-only input/output at $0.30/$2.50 per million tokens on Gemini 2.5 Flash.

2. Phase 2 — Execution. Based on the tool call, the backend executes the appropriate action. If the tool is generate_image, a separate call to the image generation model (Gemini 2.5 Flash Image or 3.1 Flash Image) produces the image at $0.039-$0.067 per image. If the tool is ask_followup or search_objects, the backend queries the SimplyScapes database directly — no LLM call needed.

Why two phases instead of one? Function calling and native image generation are separate capabilities in Gemini's architecture. You cannot declare tools and also request responseModalities: ["Image"] in the same call. The image generation models (Nano Banana series) handle image output; the standard models handle function calling. Separating intent classification from image generation also means you only pay for image generation when it's actually needed — follow-up questions and text answers skip the expensive image call entirely.

3.2 Mode as Hint vs Hard Constraint

The existing UI has a mode dropdown: Edit, Plant Selection, Ask, Finalize, Erase. The question is whether this mode should fully determine behavior or serve as a hint.

Recommendation: Mode as a strong hint that shapes the tool set, not a hard constraint.

• When the user selects Edit, the system prompt emphasizes visual modification and the generate_image tool is prioritized, but ask_followup remains available for ambiguous requests.
• When the user selects Ask, the system prompt emphasizes knowledge and answer_question is prioritized, but the AI can still suggest visual changes if relevant.
• When the user selects Erase, the tool set is restricted to generate_image (with erase-specific parameters) and request_markup — the AI can ask the user to mark what to erase but shouldn't answer unrelated questions.

This approach uses Gemini's allowed_function_names parameter within ANY mode to restrict the tool set per mode, while still allowing the AI latitude to clarify ambiguous requests.

3.3 Latency and Cost Implications

Based on published benchmarks and pricing:

• Intent classification call (Gemini 2.5 Flash, ~200 tokens in, ~50 tokens out): ~$0.000075 per call. At 1,000 daily interactions, ~$0.075/day.
• Image generation call (Gemini 2.5 Flash Image, 1024px): $0.039 per image.
• Text answer call (Gemini 2.5 Flash, ~500 tokens out): ~$0.0013 per response.

The intent classification adds one extra LLM round trip (~200-400ms based on Gemini Flash benchmarks). For synchronous workflows, this is acceptable — the user sees a brief loading state before the follow-up question appears or image generation starts. The cost is negligible relative to image generation.

Parallel function calling is supported — Gemini can call multiple tools in one response (e.g., search_objects and ask_followup simultaneously to present choices with a question). This is useful for the Plant Selection mode where the AI might search plants and formulate a question in one turn.

3.4 Migration Alert: Gemini 2.0 Flash Deprecation

Critical finding: Gemini 2.0 Flash and 2.0 Flash-Lite are being deprecated and will shut down June 1, 2026. If SimplyScapes' current v1 uses gemini-2.0-flash-exp for inpainting, migration to Gemini 2.5 Flash Image (or newer) must be part of the v2 work. The 2.5 Flash Image model costs $0.039/image at 1024px and supports the same text+image input/image output workflow.

3.5 Key Decisions for SimplyScapes

1. Two-phase routing is the recommended architecture. Cheap text model for intent classification → expensive image model only when needed.
2. Use ANY mode with allowed_function_names per UI mode to constrain the tool set while preserving flexibility.
3. Keep tool declarations to 6-8 total — well within Gemini's recommended ceiling of 10-20 for optimal accuracy.
4. Low temperature (0-0.2) for intent classification ensures deterministic routing decisions.
5. Migrate off Gemini 2.0 Flash before June 2026 — target Gemini 2.5 Flash Image or evaluate 3.1 Flash Image (faster, higher resolution, slightly more expensive).

---

4. Rich Interaction Patterns

Moving from single-turn prompting to a conversational interface requires defining how each mode behaves, when follow-ups occur, how the object library integrates, and how markup/annotation works as input.

4.1 Interaction Model Patterns from Industry Leaders

ChatGPT's image editing model sets the benchmark for multi-turn conversational image editing. GPT Image 1.5 (late 2025) introduced the ability to make precise edits while keeping existing details intact — users describe changes in natural language, and the system modifies the image iteratively without regenerating from scratch. The key pattern: edits are additive and conversational. Users say "move the tree to the left" or "make the sky warmer" and the model maintains context across turns. This is exactly the workflow SimplyScapes needs.

Canva Magic Studio takes a different approach — unified but modal. Users choose between "Design", "Image", "Doc", "Code", or "Video clip" up front (similar to SimplyScapes' mode dropdown), then interact within that mode. Canva's AI assistant supports natural language and voice commands, and can be invoked via @mention in comments. The key insight: Canva keeps the interface simple by limiting AI actions per context rather than exposing all capabilities at once. This validates SimplyScapes' mode dropdown approach.

Figma AI demonstrates in-canvas AI assistance where actions happen directly on the design surface rather than in a separate chat panel. Figma's "Code to Canvas" feature (Anthropic partnership, Feb 2026) converts AI-generated code into editable Figma frames — real design objects, not flat images. The transferable pattern: AI outputs should be native design objects when possible, not just rendered images. For SimplyScapes, this means when the AI suggests a plant, the result should ideally be a placed design object, not just a generated image with the plant painted in.

4.2 End-to-End Flows for Each Mode

Based on the two-phase routing architecture (Topic 1) and industry patterns:

Edit Mode (with follow-up capability):

User: "Add a water feature near the patio" [+ design image]
  ↓
Phase 1 (intent classification):
  Gemini analyzes prompt → detects ambiguity ("water feature" is broad)
  → calls ask_followup(question, options[]) with search_objects(category="water_features")
  ↓
Backend: Queries design_object table WHERE type = 'water_feature'
  Returns: [{name: "Tiered Fountain", thumbnail: url, id: uuid}, ...]
  ↓
Response: {type: "choices", question: "Which style?", options: [...]}
  ↓
User: Selects "Tiered Fountain"
  ↓
Phase 1 again: Gemini sees selection → calls request_markup(instruction)
  ↓
Response: {type: "markup_request", instruction: "Draw where you'd like the fountain placed"}
  ↓
User: Draws circle on design image [markup mode]
  ↓
Client sends: parts[text + selection + image + markup]
  ↓
Phase 2 (image generation):
  Gemini Image model receives: original image + markup overlay + prompt
  "Add a Tiered Fountain at the marked location"
  → Generates new design image
  ↓
Response: {type: "image", data: base64, creditsUsed: 10}

Edit Mode (direct generation — no ambiguity):

User: "Make the sky more dramatic with sunset colors" [+ design image]
  ↓
Phase 1: Gemini analyzes → clear intent, no ambiguity
  → calls generate_image(prompt, style_guidance)
  ↓
Phase 2: Image generation model produces result
  ↓
Response: {type: "image", creditsUsed: 10}

Plant Selection Mode:

User: "I need shade trees for zone 9" [+ design image]
  ↓
Phase 1: Gemini → calls search_plants(query="shade trees", filters={zone: 9})
  ↓
Backend: Queries plant table WHERE zone_min <= 9 AND zone_max >= 9
  AND sun IN ('full shade', 'partial shade') AND habit = 'tree'
  Returns top matches with images and metadata
  ↓
Response: {type: "choices", question: "Here are shade trees for zone 9:",
  options: [{name: "Crape Myrtle", thumbnail: url, ...}, ...]}
  ↓
User: Selects "Crape Myrtle"
  ↓
Phase 2: Image generation with plant placement
  ↓
Response: {type: "image", creditsUsed: 10}

Ask Mode:

User: "What ground cover works well under these oaks?"
  ↓
Phase 1: Gemini → calls answer_question(response_text)
  ↓
Response: {type: "text", text: "For shade under oaks, consider...",
  creditsUsed: 1}

Finalize Mode:

User: "Polish this design — enhance lighting and add depth" [+ design image]
  ↓
Phase 1: Gemini → calls generate_image(prompt, style_guidance="finalize")
  ↓
Phase 2: Image generation with enhancement prompt
  ↓
Response: {type: "image", creditsUsed: 10}

Erase Mode:

User: "Remove the old hedge along the fence" [+ design image]
  ↓
Phase 1: Gemini → needs to know what to erase
  → calls request_markup(instruction="Mark the hedge you want removed")
  ↓
User: Draws over the hedge [markup mode]
  ↓
Client sends: parts[text + image + markup]
  ↓
Phase 2: Image generation with mask-based inpainting
  ↓
Response: {type: "image", creditsUsed: 10}

4.3 Object Library Integration

The SimplyScapes platform has rich structured data that most AI image generators lack:

• design_object table: Hardscape items (fountains, benches, pavers, pergolas) with images, transparent cutouts, types, and tags
• plant table: 2,500+ plants with taxonomy, physical attributes, zone data, seasonal characteristics, and multiple image types
• design_material table: Texture fills (grass, mulch, stone) with type categorization

How choices are presented:

When the AI calls search_objects or search_plants, the backend queries the appropriate table and returns structured results. The chat UI renders these as selectable cards with:

• Thumbnail image (from plant_image or design_object transparent cutout)
• Name and key attributes (height, width, zone, water needs for plants)
• Category badge (e.g., "Water Feature", "Shade Tree")

Filtering strategy:

The AI's tool call includes parameters that map to database queries:

• search_objects(category, query, limit) → SELECT * FROM design_object WHERE type_id = ? AND (name ILIKE ? OR tags @> ?)
• search_plants(query, filters, limit) → SELECT * FROM plant WHERE zone_min <= ? AND zone_max >= ? AND sun = ? AND ...

This grounds the AI's suggestions in real inventory — users can only select items that actually exist in the SimplyScapes library, preventing hallucination.

4.4 Markup/Annotation as Input

User markup is a distinct input type that guides spatial placement in image generation. The implementation:

Markup mode activation: When the AI calls request_markup, the client transitions to a lightweight drawing mode. The user draws directly on the design image — freehand circles, arrows, or highlighted regions.

Markup capture: The drawing is captured as a separate image layer (PNG with transparency). Only the user's strokes are captured, not the underlying design.

Markup as a part: The markup image is sent as {type: "markup", mimeType: "image/png", data: base64} alongside the original design image. The two images are composited or sent as separate inputs to the image generation model.

How Gemini uses markup: The image generation model receives the original image and a prompt that references the markup: "Add a Tiered Fountain at the location marked by the user's drawing." The markup provides spatial guidance that text alone cannot convey — this is a significant UX advantage over pure text prompting.

Existing precedent: Mask-based inpainting is well-established — white pixels indicate the area to modify, black pixels indicate areas to preserve. SimplyScapes' markup follows this pattern but is more expressive: instead of binary mask, the user's drawing conveys intent (circled area = "here", arrow = "direction", scribble = "remove this"). The system prompt guides Gemini to interpret these cues.

4.5 Multi-Turn Context

Within a design session, conversations should maintain context across turns:

• Turn 1: "Add some Mediterranean landscaping elements"
• Turn 2: (after viewing choices) "The lavender, and also add a gravel path"
• Turn 3: "Actually, make the path curved instead of straight"

The system needs to track: (1) what was discussed, (2) what was selected, (3) what was generated. This is covered in detail in Topic 6 (Conversation Persistence).

Recommended context window: Keep the last 5-8 turns of conversation history. Beyond that, the context becomes expensive to replay and the user likely started a new train of thought. This is consistent with ChatGPT's approach where long conversations gradually lose coherence on early context.

---

5. System Prompt Prototyping

Draft system instructions and tool declarations for each mode. These are starting points for implementation — they'll need iteration based on testing with real design images.

5.1 Shared Preamble (All Modes)

All modes receive this shared context at the start of the system instruction:

You are an AI design assistant for SimplyScapes, a web-based landscape
design platform. You help landscaping professionals and homeowners
create beautiful outdoor spaces.

You are working inside the Visual Designer, where users overlay plants,
hardscape objects, and materials onto property photos.

IMPORTANT RULES:
- Only suggest plants and objects that exist in the SimplyScapes library.
  Use the search_plants and search_objects tools to find real items.
- Never hallucinate plant names, object types, or materials.
- When unsure about what the user wants, ask a clarifying question
  using the ask_followup tool rather than guessing.
- Keep responses concise — users are often on mobile devices in the field.
- Reference the design image when describing placement or changes.

5.2 Mode-Specific System Instructions

Edit Mode:

MODE: EDIT — Visual design modification

You help users modify their landscape design. Users describe changes
they want — adding elements, changing colors, adjusting the scene,
modifying backgrounds, or enhancing areas.

WHEN TO GENERATE vs CLARIFY:
- If the request is specific enough to act on (e.g., "make the sky
  bluer", "add more contrast"), use generate_image immediately.
- If the request is ambiguous or has multiple possible interpretations
  (e.g., "add a water feature", "make it look nicer"), use ask_followup
  with choices from the object library via search_objects.
- If the user needs to show you WHERE to make a change, use
  request_markup to ask them to draw on the image.

GUARDRAILS:
- Do not remove elements unless explicitly asked.
- Preserve existing plants and objects when modifying backgrounds.
- Maintain the overall style and lighting of the original photo.

Plant Selection Mode:

MODE: PLANT SELECTION — Find and place plants

You help users find the right plants for their landscape design.
Use the search_plants tool to find plants that match their criteria.

WORKFLOW:
1. Understand what the user needs (shade/sun, size, type, zone, style)
2. Search the plant library using search_plants with appropriate filters
3. Present results using ask_followup with plant options
4. When the user selects a plant, use generate_image to place it

ALWAYS search the library before suggesting plants. Never recommend
a plant by name without first confirming it exists in the
SimplyScapes plant database via search_plants.

If the user's zone, sun, or soil conditions aren't clear from
context, ask — incorrect plant recommendations are worse than
asking a question.

Ask Mode:

MODE: ASK — Knowledge and advice

You answer questions about landscaping, plants, design principles,
materials, and outdoor spaces. You are knowledgeable but not a
replacement for a licensed landscape architect.

BEHAVIOR:
- Provide helpful, accurate answers about landscaping topics.
- When referencing specific plants, use search_plants to verify
  they exist in the SimplyScapes library and include relevant
  attributes (zone range, water needs, size).
- If a question would be better served by a visual change,
  suggest the user switch to Edit or Plant Selection mode.
- Keep answers concise (2-3 paragraphs max).

Do NOT generate images in this mode. Use answer_question for
all responses.

Finalize Mode:

MODE: FINALIZE — Polish and enhance

You enhance the overall quality of the design image for
presentation purposes — improving lighting, adding depth,
enhancing colors, and making the scene look professionally
rendered.

BEHAVIOR:
- Apply photorealistic enhancement to the entire image.
- Maintain all existing design elements (plants, objects,
  structures) exactly as placed.
- Focus on: lighting consistency, shadow depth, color
  vibrancy, atmospheric quality, and professional polish.
- If the user provides specific enhancement requests,
  prioritize those.

GUARDRAILS:
- Never add or remove design elements.
- Never change plant species or object types.
- Never alter the fundamental composition or layout.

Use generate_image with style_guidance="finalize" for all
requests in this mode.

Erase Mode:

MODE: ERASE — Remove elements from the design

You help users remove unwanted elements from their landscape
design — old plants, structures, objects, or background elements.

WORKFLOW:
1. If the user describes WHAT to remove but not WHERE, use
   request_markup to ask them to mark the area.
2. If the user provides both description and markup, use
   generate_image with the mask to remove the element and
   fill with appropriate background.

GUARDRAILS:
- Only remove what the user explicitly asks to remove.
- Fill erased areas with contextually appropriate background
  (match surrounding grass, fence, sky, etc.).
- Never add new elements during an erase operation.

Do NOT use ask_followup in this mode unless the user's
request is truly incomprehensible. Erase should feel fast
and direct.

5.3 Tool Declarations (JSON Schema)

These are the Gemini function calling tool declarations, shared across modes with per-mode filtering via allowed_function_names:

{
  "tools": [{
    "function_declarations": [
      {
        "name": "generate_image",
        "description": "Generate or modify the design image based on the user's request. Call this when you have enough information to produce a visual result.",
        "parameters": {
          "type": "object",
          "properties": {
            "prompt": {
              "type": "string",
              "description": "Detailed prompt describing the desired image modification. Be specific about what to add, change, or enhance."
            },
            "style_guidance": {
              "type": "string",
              "enum": ["edit", "finalize", "erase", "plant_placement"],
              "description": "The type of image operation to perform."
            },
            "placement_hints": {
              "type": "string",
              "description": "Optional spatial guidance for element placement (e.g., 'left side of patio', 'along the fence line'). Use when the user has described a location in text."
            }
          },
          "required": ["prompt", "style_guidance"]
        }
      },
      {
        "name": "ask_followup",
        "description": "Ask the user a clarifying question, optionally presenting choices from the object or plant library. Use when the request is ambiguous or multiple options exist.",
        "parameters": {
          "type": "object",
          "properties": {
            "question": {
              "type": "string",
              "description": "The question to ask the user. Keep it concise and actionable."
            },
            "options": {
              "type": "array",
              "description": "Optional list of choices to present. Each option references a SimplyScapes library item.",
              "items": {
                "type": "object",
                "properties": {
                  "id": { "type": "string", "description": "The SimplyScapes object or plant ID" },
                  "name": { "type": "string", "description": "Display name" },
                  "type": { "type": "string", "enum": ["design_object", "plant", "design_material"] }
                },
                "required": ["id", "name", "type"]
              }
            }
          },
          "required": ["question"]
        }
      },
      {
        "name": "search_objects",
        "description": "Search the SimplyScapes design object library for hardscape items like fountains, benches, pavers, pergolas, fire pits, etc.",
        "parameters": {
          "type": "object",
          "properties": {
            "category": { "type": "string", "description": "Object category to filter by (e.g., 'water_feature', 'seating', 'paving', 'lighting', 'structure')" },
            "query": { "type": "string", "description": "Free-text search query" },
            "limit": { "type": "integer", "description": "Max results to return (default 5, max 10)" }
          },
          "required": ["query"]
        }
      },
      {
        "name": "search_plants",
        "description": "Search the SimplyScapes plant library. Supports filtering by USDA zone, sun exposure, water needs, plant type, and size.",
        "parameters": {
          "type": "object",
          "properties": {
            "query": { "type": "string", "description": "Free-text search (common name, genus, species)" },
            "zone": { "type": "integer", "description": "USDA hardiness zone (1-13)" },
            "sun": { "type": "string", "enum": ["full_sun", "partial_sun", "partial_shade", "full_shade"] },
            "water": { "type": "string", "enum": ["low", "moderate", "high"] },
            "plant_type": { "type": "string", "description": "e.g., 'tree', 'shrub', 'perennial', 'annual', 'ground_cover', 'grass'" },
            "limit": { "type": "integer", "description": "Max results (default 5, max 10)" }
          },
          "required": ["query"]
        }
      },
      {
        "name": "answer_question",
        "description": "Provide a text-only answer to a knowledge question about landscaping, plants, design, or materials. No image generation.",
        "parameters": {
          "type": "object",
          "properties": {
            "response_text": {
              "type": "string",
              "description": "The answer text. Keep concise (2-3 paragraphs max)."
            }
          },
          "required": ["response_text"]
        }
      },
      {
        "name": "request_markup",
        "description": "Ask the user to draw/annotate on the design image to indicate a location or area. The client will enter markup mode.",
        "parameters": {
          "type": "object",
          "properties": {
            "instruction": {
              "type": "string",
              "description": "Tell the user what to mark. Be specific: 'Circle the area where you want the fountain' or 'Draw over the hedge you want removed'."
            }
          },
          "required": ["instruction"]
        }
      }
    ]
  }]
}

5.4 Tool Availability Per Mode

| Tool | Edit | Plant Selection | Ask | Finalize | Erase | |------|------|-----------------|-----|----------|-------| | generate_image | Yes | Yes | No | Yes | Yes | | ask_followup | Yes | Yes | No | No | No | | search_objects | Yes | No | Yes | No | No | | search_plants | No | Yes | Yes | No | No | | answer_question | No | No | Yes | No | No | | request_markup | Yes | No | No | No | Yes |

This matrix is enforced via allowed_function_names in the Gemini API call. For example, Edit mode sends:

"tool_config": {
  "function_calling_config": {
    "mode": "ANY",
    "allowed_function_names": ["generate_image", "ask_followup", "search_objects", "request_markup"]
  }
}

5.5 Prompt Engineering Notes

Based on Gemini best practices research:

1. Keep system instructions concise. Gemini 3+ models reason naturally and perform worse with over-explained prompts. The mode-specific instructions above are intentionally short.
2. Place instructions before input. Gemini performs better when format and rules come first, followed by the user's content.
3. Use low temperature (0-0.2) for intent classification. This ensures deterministic tool selection. The image generation model can use higher temperature for creative output.
4. Limit to 6 tools total. Within the recommended ceiling of 10-20, but keeping it tight improves selection accuracy.
5. Enum constraints on parameters. style_guidance, sun, water, plant_type use enums to prevent invalid values and improve parameter accuracy.

---

6. Credit System Architecture

The credit system must balance three concerns: (1) covering actual Gemini API costs with healthy margin, (2) creating predictable pricing that users understand, and (3) providing enough granularity to differentiate cheap operations (text answers) from expensive ones (image generation).

6.1 Competitive Pricing Landscape

The adjacent market analysis reveals clear pricing patterns across seven AI-powered creative platforms:

| Product | Model | Free Tier | Entry Paid | Credit Unit | Overage | |---------|-------|-----------|------------|-------------|---------| | Canva AI | Bundled subscription | 50 total (lifetime) | $15/mo → 500/mo | Per generation | Hard wall | | Figma AI | Per-seat credits | 500/mo (Starter) | $5/editor/mo → 3,000/mo | Variable by action (30-100+) | Enforcement Mar 2026 | | Adobe Firefly | Tiered credits | Via free CC | $9.99/mo → 2,000/mo | 1/image; 20-100/sec video | On-demand purchase | | Midjourney | GPU time | None | $10/mo → 3.3 GPU hrs | GPU minutes | Relax Mode (unlimited, lower priority) | | Runway ML | Credits + unlimited | 125 one-time | $12/mo → 625/mo | $0.01/credit | Buy more or Relax Mode | | ChatGPT | Rate-limited | 2-3 images/day | $20/mo → 50/3hrs | Images per window | Wait for reset | | LeanScaper | Credits (landscaping) | 250/mo | $300/mo → 3,000/mo | Per AI action | $150/1,000 credits top-up |

Key patterns:

• Credits are dominant — 5 of 7 products use credit systems. Time-window rate limiting (ChatGPT) is the exception.
• Variable cost by complexity — Figma charges 30-100+ credits depending on action complexity. Adobe charges 1 credit for images, 20-100/sec for video. This is best practice.
• "Unlimited at lower priority" reduces anxiety — Midjourney and Runway both offer unlimited generation at relaxed priority. This is powerful for creative exploration.
• Free tiers are tight but present — Canva (50 lifetime), ChatGPT (2-3/day), Runway (125 one-time). Enough to evaluate, not enough to work.
• Team credit pooling is an enterprise upsell — Adobe and Figma both offer or are introducing shared pools.

In the vertical market, LeanScaper is the only landscape competitor with a credit system — $300-$1,500/mo with 3,000-18,000 credits and $150/1,000 top-ups. Their credits cover business operations (marketing, financials, SOPs), not design generation.

6.2 Cost Foundation: Gemini API Pricing

The credit system must be anchored to actual costs. Current Gemini pricing (March 2026):

| Model | Use Case | Cost | Notes | |-------|----------|------|-------| | Gemini 2.5 Flash (text) | Intent classification | $0.30 input / $2.50 output per 1M tokens | ~$0.00008 per classification call | | Gemini 2.5 Flash Image | Image generation | $0.039 per image (1024px) | Primary image model | | Gemini 3.1 Flash Image | Image generation | $0.067 per image (1024px) | Faster, higher res, more expensive | | Gemini 2.5 Flash (text) | Text answer | ~$0.0013 per response (500 tokens) | Ask mode answers |

Per-interaction cost breakdown:

| Operation | Gemini Calls | Estimated Cost | Credit Cost | Margin | |-----------|-------------|---------------|-------------|--------| | Image generation (edit/finalize/erase) | 1 text + 1 image | ~$0.040 | 10 credits | ~100x at $0.40/10 credits | | Image gen with follow-up first | 2 text + 1 image | ~$0.040 | 10 credits (gen only) | Same — follow-up is free | | Text answer (ask mode) | 1 text | ~$0.001 | 2 credits | ~100x at $0.08/2 credits | | Follow-up question | 1 text | ~$0.00008 | 0 credits | Subsidized | | Object/plant search | 0 (DB query) | ~$0.00 | 0 credits | Pure DB cost |

The high margin is intentional — it covers infrastructure overhead, rate limiting, conversation state storage, credit system management, and provides room for Gemini pricing increases. At the recommended credit price of ~$0.04/credit, margins are healthy even if Gemini costs increase 5-10x.

6.3 Credit Cost Table

| Operation | Credits | Rationale | |-----------|---------|-----------| | Image generation (edit, finalize, erase) | 10 | Full Gemini image model call — the expensive operation | | Image generation with markup | 10 | Same cost — markup is additional input, not additional inference | | Plant placement generation | 10 | Image generation with plant overlay | | Text response (ask mode) | 2 | Text-only LLM call, significantly cheaper | | Follow-up / clarification question | 0 | Encourages exploration; negligible LLM cost for intent classification | | Object library search | 0 | Local database query, no LLM cost | | Plant library search | 0 | Local database query, no LLM cost | | Markup request | 0 | No LLM cost — client-side annotation mode |

Design principles:

• Free actions encourage engagement. Making follow-ups and searches free removes friction from the conversational flow. Users shouldn't hesitate to clarify or browse.
• 10-credit generations align with the user's mental model. "10 credits = 1 image" is easy to understand and remember.
• Text answers are cheap but not free. 2 credits acknowledges LLM cost while keeping Ask mode accessible. At 2 credits, a user with 100 credits gets 10 images or 50 text answers.

6.4 Tier Structure

Based on competitive analysis and the cost model:

| Tier | Monthly Credits | Image Generations | Monthly Price | Per-Credit Cost | Target User | |------|----------------|-------------------|---------------|-----------------|-------------| | Free | 50 | ~5 images | $0 | N/A | Homeowner evaluation | | Starter (subscriber) | 200 | ~20 images | Included with subscription | Bundled | Active subscriber | | Pro (subscriber) | 500 | ~50 images | Included with subscription | Bundled | Professional subscriber | | Credit Pack (add-on) | 500 | ~50 images | $19.99 | $0.04/credit | Top-up for any tier | | Credit Pack (add-on) | 1,500 | ~150 images | $49.99 | $0.033/credit | Volume top-up |

Design decisions:

1. Free tier: 50 credits/month (not lifetime). Canva's 50-lifetime cap is stingy. Monthly reset encourages repeat visits. 50 credits = 5 images + some text answers — enough to experience the feature, not enough for production work.

2. Bundled with subscription. AI credits are included in the SimplyScapes subscription tiers (Starter, Pro), not sold separately as a base product. This follows Canva and Adobe's model of bundling AI into existing plans.

3. Credit packs as add-ons. Subscribers who need more credits can purchase packs. This follows Adobe's on-demand purchase model and LeanScaper's $150/1,000 top-up pattern.

4. Workspace-scoped, not per-user. Credits belong to the workspace. All team members draw from the same pool. This is simpler than per-user allocation and matches SimplyScapes' workspace-first architecture.

5. No "unlimited" tier initially. The Midjourney/Runway "relaxed unlimited" pattern is attractive but adds infrastructure complexity (priority queuing). Consider for v2.

6.5 Workspace Credit Management

Who manages credits:

• Workspace admins can view credit balance, purchase credit packs, and see usage history
• All workspace members can use credits (no per-user limits initially)
• Credit balance is displayed in the AI tab header (already implemented per PR #1166)

Credit deduction flow:

1. User sends request → API receives
2. API checks workspace credit balance
3. If insufficient credits → return {type: "error", code: "insufficient_credits"}
4. Pre-deduct credits (optimistic deduction)
5. Call Gemini API
6. If Gemini fails → refund credits, return error
7. If Gemini succeeds → confirm deduction, return response

Race condition handling: Use Hasura's atomic operations (e.g., update_workspace SET credits = credits - 10 WHERE credits >= 10 RETURNING credits). The WHERE credits >= 10 clause prevents negative balances without explicit locking.

Credit exhaustion mid-conversation: When credits run out during a conversation, the current turn completes (credits were pre-deducted), but the next generation request returns an insufficient_credits error with a prompt to purchase more. Follow-ups and searches remain free, so the conversation can continue — the user just can't generate images.

6.6 Whitelabel Credit Model

For whitelabel instances (the instance system):

• Each instance has independent credit configuration. The whitelabel partner decides their own credit allocation per tier.
• SimplyScapes bills the partner at wholesale. Partners purchase credits in bulk and resell at their own markup.
• Per-instance enable/disable. Already a planned task — the API checks instance.ai_enabled before processing requests.
• Instance-scoped credit pools. Credits are tracked per workspace within the instance. The partner's admin dashboard shows aggregate usage.

6.7 Technical Implementation: Stripe

SimplyScapes already has Stripe billing. For credits:

Stripe is the recommended path for credit packs. Stripe's usage-based billing now supports credit-based pricing natively — customers prepay for credits that are spent in real time. Key capabilities:

• Meters for tracking credit consumption events
• Billing credits that can be granted, spent, and tracked
• Recurring credit grants (monthly allocation with subscription)
• Proration and rollover policy control
• Up to 100M usage events/month included with Stripe Billing

Implementation approach:

1. Subscription tiers (Stripe, existing): Include monthly credit allocation as a plan feature
2. Credit pack purchases (Stripe): One-time checkout session for credit packs
3. Credit tracking (Hasura): workspace_ai_credits table tracking balance, transactions, and usage history
4. Real-time deduction (API): Atomic credit deduction on each AI operation

Hasura schema sketch:

-- Credit balance per workspace
CREATE TABLE workspace_ai_credits (
  workspace_id UUID PRIMARY KEY REFERENCES workspace(id),
  balance INTEGER NOT NULL DEFAULT 0,
  monthly_allocation INTEGER NOT NULL DEFAULT 0,
  last_reset_at TIMESTAMPTZ,
  updated_at TIMESTAMPTZ DEFAULT now()
);

-- Credit transaction log
CREATE TABLE ai_credit_transaction (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  workspace_id UUID NOT NULL REFERENCES workspace(id),
  amount INTEGER NOT NULL, -- positive = credit, negative = debit
  type TEXT NOT NULL, -- 'monthly_grant', 'pack_purchase', 'generation', 'text_answer', 'refund'
  reference_id TEXT, -- Stripe payment ID, AI request ID, etc.
  created_at TIMESTAMPTZ DEFAULT now(),
  created_by UUID REFERENCES "user"(id)
);

---

7. API Endpoint Design

The API endpoint is the contract between the designer client and the backend AI system. It must handle multi-part payloads (text, images, masks, markup, selections), route to the correct Gemini model based on intent classification, manage conversation state, enforce credit limits, and return polymorphic responses (images, text, choices, errors).

7.1 Endpoint Specification

POST /api/v1/ai/generate
Content-Type: application/json
Authorization: Bearer <Firebase JWT>

Request:

type AIMode = "edit" | "plant_selection" | "ask" | "finalize" | "erase";

type Part =
  | { type: "text"; content: string }
  | { type: "image"; mimeType: string; data: string }       // Design image (base64)
  | { type: "mask"; mimeType: string; data: string }         // Binary mask for erase
  | { type: "markup"; mimeType: string; data: string }       // User annotation overlay
  | { type: "selection"; objectId: string; objectType: string; label?: string };

interface AIGenerateRequest {
  mode: AIMode;
  parts: Part[];
  designId: string;
  conversationId?: string;    // Omit for new conversation
  userId: string;             // From Firebase JWT (validated server-side)
  workspaceId: string;        // From JWT claims
  instanceId?: string;        // Whitelabel instance
}

Response:

interface AIGenerateResponse {
  id: string;                 // Unique response ID
  conversationId: string;     // Created on first turn, returned for subsequent turns
  type: "image" | "text" | "choices" | "markup_request" | "error";

  // Present when type = "image"
  image?: {
    data: string;             // Base64 encoded
    mimeType: string;
    promptUsed: string;       // The actual prompt sent to Gemini (for debugging/transparency)
  };

  // Present when type = "text"
  text?: string;

  // Present when type = "choices"
  choices?: {
    question: string;
    options: Array<{
      id: string;
      name: string;
      thumbnail?: string;     // URL to SimplyScapes CDN
      type: "design_object" | "plant" | "design_material";
      description?: string;
      metadata?: Record<string, unknown>;  // Zone, height, sun, etc.
    }>;
  };

  // Present when type = "markup_request"
  markupRequest?: {
    instruction: string;      // "Circle where you want the fountain placed"
  };

  // Present when type = "error"
  error?: {
    code: "insufficient_credits" | "rate_limited" | "content_filtered"
        | "generation_failed" | "instance_disabled" | "invalid_request";
    message: string;
    retryable: boolean;
  };

  // Always present
  usage: {
    creditsUsed: number;
    creditsRemaining: number;
    model: string;            // e.g. "gemini-2.5-flash-image"
  };
}

Changes from brainstorm draft: Added markup_request response type (separate from choices), added retryable to errors, added metadata to choice options for plant attributes, added promptUsed for transparency.

7.2 Request Processing Pipeline

Client Request
  │
  ├─ 1. Auth: Validate Firebase JWT, extract userId + workspaceId
  ├─ 2. Instance Check: If instanceId, verify ai_enabled = true
  ├─ 3. Credit Check: Verify workspace has sufficient credits
  ├─ 4. Rate Limit: Check per-workspace rate limit (e.g., 30 req/min)
  ├─ 5. Conversation Load: If conversationId, load last N turns
  │
  ├─ 6. Phase 1 — Intent Classification
  │     ├─ Build system prompt (shared preamble + mode-specific)
  │     ├─ Include tool declarations (filtered by mode)
  │     ├─ Include conversation history (if any)
  │     ├─ Send text parts to Gemini 2.5 Flash (text model)
  │     └─ Receive: function call (tool name + args) OR text response
  │
  ├─ 7. Tool Execution
  │     ├─ generate_image → Phase 2 (image generation)
  │     ├─ ask_followup → Build choices response from args
  │     ├─ search_objects → Query design_object table → format as choices
  │     ├─ search_plants → Query plant table → format as choices
  │     ├─ answer_question → Extract text → build text response
  │     └─ request_markup → Build markup_request response
  │
  ├─ 8. Phase 2 — Image Generation (only if generate_image called)
  │     ├─ Build image generation prompt from tool args
  │     ├─ Include: original design image + mask/markup (if provided)
  │     ├─ Call Gemini 2.5 Flash Image (or 3.1 Flash Image)
  │     ├─ Set responseModalities: ["TEXT", "IMAGE"]
  │     └─ Extract generated image from response
  │
  ├─ 9. Credit Deduction: Deduct credits based on operation type
  ├─ 10. Conversation Save: Store turn in conversation history
  └─ 11. Response: Return AIGenerateResponse

7.3 Gemini Translation Layer

The translation layer converts SimplyScapes' request format to Gemini's API format. This is the abstraction boundary that enables future provider swapping.

// Provider interface — Gemini-first, but swappable
interface AIProvider {
  classifyIntent(
    systemPrompt: string,
    tools: ToolDeclaration[],
    parts: GeminiPart[],
    history: ConversationTurn[],
    config: { temperature: number; toolConfig: ToolConfig }
  ): Promise<IntentResult>;

  generateImage(
    prompt: string,
    referenceImage: Buffer,
    mask?: Buffer,
    config: { model: string; size: number }
  ): Promise<GeneratedImage>;
}

// Gemini-specific implementation
class GeminiProvider implements AIProvider {
  async classifyIntent(/* ... */): Promise<IntentResult> {
    const response = await genAI.models.generateContent({
      model: "gemini-2.5-flash",
      contents: this.buildContents(parts, history),
      config: {
        systemInstruction: systemPrompt,
        temperature: 0.1,
        tools: [{ functionDeclarations: tools }],
        toolConfig: { functionCallingConfig: config.toolConfig },
      },
    });
    return this.parseIntentResult(response);
  }

  async generateImage(/* ... */): Promise<GeneratedImage> {
    const response = await genAI.models.generateContent({
      model: "gemini-2.5-flash-image",
      contents: this.buildImageContents(prompt, referenceImage, mask),
      config: {
        responseModalities: ["TEXT", "IMAGE"],
      },
    });
    return this.extractImage(response);
  }
}

Key abstraction decisions:

• classifyIntent and generateImage are separate methods — reflecting the fundamental Gemini constraint that function calling and image generation can't be in the same API call.
• The provider interface is minimal — only the methods SimplyScapes actually needs. Not a full LLM abstraction.
• Parts translation (SimplyScapes format → Gemini format) happens inside the provider, not in the API route.

7.4 App Router vs Pages Router

The current endpoint is at src/pages/api/ai-inpanting.ts (Pages Router). The recommendation:

Migrate to App Router Route Handler at src/app/api/v1/ai/generate/route.ts.

Reasons:

• App Router is Next.js' future — Pages Router is in maintenance mode
• Route Handlers support streaming responses natively via ReadableStream (for future text streaming)
• Server Actions pattern is cleaner for auth validation
• Colocated with the rest of the v2 AI infrastructure
• The existing Pages Router endpoint continues working during migration — no breaking change

Migration path:

1. Build the new endpoint at src/app/api/v1/ai/generate/route.ts
2. Keep src/pages/api/ai-inpanting.ts running for v1 compatibility
3. Designer client switches to the new endpoint when v2 chat UI ships
4. Deprecate the old endpoint after v2 is stable

7.5 Validation and Rate Limiting

Request validation:

• mode must be one of the 5 valid modes
• parts must contain at least one text part
• Image parts must be valid base64 with accepted MIME types (image/png, image/jpeg, image/webp)
• designId must reference an existing design in the user's workspace
• userId is validated against the JWT (not trusted from the request body)

Rate limiting:

• Per-workspace: 30 requests/minute (prevents abuse, allows burst usage during design sessions)
• Per-user within workspace: No limit initially (workspace credits are the natural throttle)
• Global: Gemini API rate limits are the backstop

Implementation: Use a simple Redis-based or in-memory rate limiter. Vercel's edge middleware can handle this, or use a lightweight library like @upstash/ratelimit.

7.6 Error Handling Matrix

| Error | Code | HTTP Status | Retryable | User-Facing Message | |-------|------|-------------|-----------|---------------------| | Workspace has no credits | insufficient_credits | 402 | No | "You're out of AI credits. Purchase more in Settings." | | Rate limit exceeded | rate_limited | 429 | Yes | "Slow down — try again in a moment." | | Gemini content filter | content_filtered | 200 | No | "The AI couldn't process that request. Try rephrasing." | | Gemini API error/timeout | generation_failed | 200 | Yes | "Generation failed. Trying again..." | | AI disabled for instance | instance_disabled | 403 | No | "AI features are not available." | | Invalid request format | invalid_request | 400 | No | "Something went wrong. Please try again." |

Note: Content filter and generation failure return HTTP 200 with an error response body (not 4xx/5xx). This is intentional — the API processed the request successfully; the AI model declined or failed. The client distinguishes via response.type === "error".

7.7 Synchronous-First, Streaming-Ready

The v2 endpoint is synchronous — the client sends a request and waits for the complete response. This is simpler to implement and sufficient for image generation (which returns a single image, not a token stream).

Streaming upgrade path for text responses: When Ask mode text responses become long enough to benefit from streaming, the endpoint can:

1. Accept an Accept: text/event-stream header
2. Return SSE-formatted chunks for text responses
3. Continue returning JSON for image/choices/error responses

This can be implemented using Next.js Route Handler's native ReadableStream support or the Vercel AI SDK's streamText utility. The API contract remains backward-compatible — clients that don't send the streaming header get the synchronous response.

---

8. Conversation Persistence & History

Conversation state enables multi-turn interactions — "add a water feature" → (selects fountain) → "actually, put it closer to the patio" — without losing context between turns.

8.1 Storage Strategy: Server-Side (Hasura)

Recommendation: Store conversations in Hasura, not client-side.

Reasons:

• Conversations persist across sessions — user can close the designer, come back, and continue
• Analytics and usage tracking are server-side concerns
• Multi-device support — start a conversation on desktop, continue on mobile
• Audit trail for credit usage
• Foundation for future "workspace AI memory" capabilities

Alternative considered: Client-side state. Simpler to implement but conversations vanish on page reload, can't be analyzed, and don't support multi-device. Rejected for production.

8.2 Hasura Schema

-- Conversation container
CREATE TABLE ai_conversation (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  design_3d_id UUID NOT NULL REFERENCES design_3d(id) ON DELETE CASCADE,
  workspace_id UUID NOT NULL REFERENCES workspace(id),
  user_id UUID NOT NULL REFERENCES "user"(id),
  mode TEXT NOT NULL,                    -- Initial mode (may change within conversation)
  created_at TIMESTAMPTZ DEFAULT now(),
  updated_at TIMESTAMPTZ DEFAULT now()
);

-- Individual turns within a conversation
CREATE TABLE ai_conversation_turn (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  conversation_id UUID NOT NULL REFERENCES ai_conversation(id) ON DELETE CASCADE,
  turn_number INTEGER NOT NULL,
  role TEXT NOT NULL,                    -- 'user' | 'model' | 'tool_result'
  parts JSONB NOT NULL,                 -- Array of parts (text, image refs, selections)
  tool_calls JSONB,                     -- Function calls made by the model
  response_type TEXT,                   -- 'image' | 'text' | 'choices' | 'markup_request'
  credits_used INTEGER DEFAULT 0,
  model_used TEXT,                      -- 'gemini-2.5-flash', 'gemini-2.5-flash-image'
  created_at TIMESTAMPTZ DEFAULT now()
);

-- Indexes
CREATE INDEX idx_conv_design ON ai_conversation(design_3d_id);
CREATE INDEX idx_conv_workspace ON ai_conversation(workspace_id);
CREATE INDEX idx_turn_conv ON ai_conversation_turn(conversation_id, turn_number);

Relationship to existing ai_inpainting table: The new schema replaces the ai_inpainting table for v2 conversations. The existing table continues to serve v1 requests. Migration is not required — v1 and v2 coexist.

8.3 Context Replay Strategy

When a request includes a conversationId, the API loads previous turns and replays them to Gemini for context:

Sliding window: Last 5-8 turns. This balances context quality with token cost:

• 5 turns covers most follow-up flows (clarify → select → markup → generate → refine)
• 8 turns handles longer design sessions
• Beyond 8 turns, early context is usually irrelevant (the user has moved on)

Token budget: At ~500 tokens per turn average (text + tool calls), 8 turns = ~4,000 tokens of history. This is well within Gemini 2.5 Flash's 1M token context window and adds ~$0.0012 to the intent classification cost — negligible.

What's included in replayed turns:

• User text prompts
• Model tool calls and text responses
• Tool results (summarized — not full image data)
• User selections (which plant/object was chosen)

What's NOT replayed:

• Full image data from previous turns (too expensive to re-encode)
• Credit usage details
• Technical metadata

8.4 Analytics Capture

Every conversation turn creates analytics events:

| Event | Fields | Purpose | |-------|--------|---------| | ai.request | mode, hasImage, hasMarkup, workspaceId | Usage patterns by mode | | ai.intent | toolCalled, mode, classificationTime | Intent routing accuracy | | ai.generation | model, generateTime, imageSize | Performance monitoring | | ai.credits | creditsUsed, creditsRemaining, operation | Credit consumption tracking | | ai.error | errorCode, model, retried | Error rate monitoring | | ai.conversation | turnCount, totalCredits, duration | Session analysis |

These events power the admin dashboard for credit usage, the workspace usage history view, and internal product analytics.

8.5 Future: Workspace AI Memory

The conversation persistence schema is designed to support future "AI memory" capabilities:

• Cross-conversation learning: Aggregate plant preferences, style tendencies, and frequent requests across all conversations in a workspace
• Design context: When starting a new conversation on an existing design, the AI knows what was previously generated and discussed
• pgvector embeddings: Conversation summaries can be embedded and stored for semantic retrieval — "what did we discuss about the backyard last month?"

These are not v2 features, but the schema design doesn't block them.

---

Part III: Market Landscape

9. Market Overview

How the market currently handles AI in landscape design:

The landscape design market is split into two distinct camps: products deeply invested in AI (PRO Landscape+, Yardzen, LeanScaper) and products with essentially no AI capabilities (iScape, Planter). There is no middle ground — companies either committed to AI as a core strategy or haven't started. None offer what SimplyScapes is building: a conversational AI design partner.

PRO Landscape+ has the most AI features (5 tools including an outdoor living designer and AI eraser) but is Windows-only, desktop-only, and professional-only. Yardzen offers free AI-generated inspiration images but charges $995-$1,995 for human-designed plans with no self-service middle tier. LeanScaper has built a chat-based AI — but for business operations, not design. iScape has 4 million downloads but relies on AR, not AI.

In adjacent markets, the creative AI tools (Canva, Figma, Midjourney, Adobe Firefly, Runway, ChatGPT) have established mature interaction patterns for AI-assisted creation — but none have domain knowledge about plants, hardiness zones, companion planting, or landscape design principles.

Market maturity: Emerging — AI adoption is bimodal with no established middle ground. Customer satisfaction: Underserved — the gap between free AI inspiration and $995+ human design is the largest unmet need.

10. Vertical Market Analysis

Full analysis available in supporting/vertical-competitor-analysis.md. Key findings per competitor:

LeanScaper

Approach: Chat-based AI for landscape business operations (not design). Specialized agents (CFO, CMO, SOP) handle financial analysis, marketing, and process documentation. Voice interaction for fieldwork. Strengths: Validates conversational AI for the landscape vertical. Professionals engage with chat-based AI for complex tasks. Credit-based pricing works. Gaps: No design visualization, no image generation, no spatial design. Entirely non-overlapping with SimplyScapes' design focus. Takeaway: LeanScaper proves the interaction model; SimplyScapes brings it to design.

iScape

Approach: Photo-based landscape design via AR (augmented reality). Drag-and-drop placement of plants and objects on user's yard photo. 4M downloads. Strengths: Photo-based design on your own yard is the baseline expectation. Professional proposal generation bridges design to sales. Gaps: No AI features. No chat interface. Users must manually browse catalogs and place every element. App Store reviews criticize the absence of AI. Takeaway: The "design on your own photo" paradigm is validated. SimplyScapes leapfrogs by adding conversational AI to this paradigm.

PRO Landscape+ (Drafix)

Approach: Most AI-forward professional tool with 5 AI features. AI-to-CAD pipeline (concept → scaled drawing → material takeoff → proposal). Windows desktop only. Strengths: AI outputs feed directly into CAD for actionable professional output. 1,000+ real manufacturer paver patterns. Ask Wayne chatbot for software help. Gaps: Windows-only desktop. Professional-only ($900/yr). No natural language design requests. Ask Wayne is limited to software help, not design guidance. Takeaway: PRO Landscape+ is the quality bar for professional AI output, but their desktop-only, Windows-only positioning leaves the entire mobile-first and homeowner market open.

Yardzen

Approach: Hybrid AI + human design service. Free YardAI generates instant concepts from photos (16 style choices). Paid packages ($995-$1,995) add human designers. Trained on 50,000 real designs. Strengths: Free AI tool is a powerful acquisition channel. 28% conversion increase after leaning into human touchpoints over AI-only. Gaps: YardAI generates inspiration only — no interactive refinement, no actionable plans. $995 gap between free AI and human design. Takeaway: The "inspiration → professional design" gap is the core opportunity. SimplyScapes fills it with conversational AI that produces actionable output at $50-200 price points.

Planter

Approach: Focused vegetable garden planner with rule-based companion planting intelligence. Grid-based drag-and-drop. No AI (beyond icon generation). Strengths: Demonstrates that plant intelligence (companion planting, spacing, zone awareness) creates genuine user value even without AI. Gaps: No landscape design. No AI. No chat interface. Takeaway: The horticultural intelligence (zone data, companion planting, spacing) should be embedded in SimplyScapes' AI as foundational knowledge — then AI-powered generation layers on top.

Vertical market patterns:

1. AI is used for generation, not conversation. No competitor offers a conversational design partner.
2. Photo-based design is the baseline expectation. Every design competitor starts from the user's yard photo.
3. The funnel gap is universal. Free AI inspiration → $995+ human design, with nothing in between.
4. Human-AI hybrid outperforms AI-only. Yardzen's 28% conversion increase validates that consumers want AI assistance + human validation.
5. Credit-based pricing is emerging. LeanScaper's credit model validates consumption-based pricing in the landscape vertical.

11. Adjacent Market Patterns

Full analysis available in supporting/adjacent-market-analysis.md. Key transferable patterns:

Canva AI — Contextual AI placement

Pattern: AI tools are embedded at the point of need, not siloed. Progressive disclosure via tiered credit limits (50 free / 500 Pro). Multi-backend abstraction (DALL-E + Imagen behind one interface). Adaptation: Embed AI generation directly into the landscape canvas — select empty area → "AI Fill" contextually. Route landscaping prompts to best model based on task. What doesn't transfer: Template-based paradigm breaks for site-specific landscape designs requiring 3D spatial reasoning.

Figma AI — Bidirectional canvas integration

Pattern: AI output is fully editable structured objects (not flat images). Variable credit cost by action complexity (30-100+ credits). Non-prompt AI tools (erase/isolate without text input). Adaptation: When AI suggests a plant, the result should ideally be a placed design object, not just a painted image. Offer direct-manipulation AI (select area → "fill with seasonal color"). What doesn't transfer: Figma's component library assumes identical reusable elements; landscape elements are variable.

Midjourney — Conversational prompt translation

Pattern: Users speak naturally; AI translates to optimized generation prompts. 2x2 grid for comparing variations. Voice-driven ideation. Preference learning from image ratings. Adaptation: Translate "I want a cozy backyard" into specific plant species, materials, and layout. Show 4 landscape variations for comparison. Voice-driven on-site design sessions. What doesn't transfer: Generates artistic images, not actionable plans. GPU-time billing doesn't map to design tasks.

ChatGPT Image Generation — Multi-turn conversational editing

Pattern: Users edit images through natural language dialogue ("move the tree left", "make the path wider"). Context-aware across multiple turns. No modal switch between chat and generation. Adaptation: Build iterative refinement where each turn modifies the existing design, not regenerates from scratch. Maintain full conversation context for coherent changes. What doesn't transfer: Generates flat raster images with no structured data (no plant names, no dimensions, no bill of materials).

Vercel AI SDK — Provider-agnostic infrastructure

Pattern: Unified interface across OpenAI, Anthropic, Google. Type-safe tool calling with Zod schemas. Agent loop for multi-step operations. Streaming-first. Next.js-optimized. Adaptation: Use the provider abstraction to start with Gemini and swap/add providers later. Define landscape-specific tools with Zod schemas for validated AI suggestions. Consider for the SDK layer. What doesn't transfer: Infrastructure only — provides building blocks but not domain logic.

Cross-industry insights not yet applied to landscaping:

1. Conversational-to-structured-output. No product generates structured spatial designs through conversation — every tool produces either flat images or structured UI objects, never structured landscape plans.
2. Passive preference learning. Midjourney's image-rating personalization hasn't been applied — show 30 landscape photos at onboarding, learn aesthetic preferences.
3. Tiered quality/cost generation. Quick sketches, detailed plans, and photorealistic presentations should be different products at different price points.
4. Enterprise credit pooling. Landscape companies with 3-15 designers need shared pools, not per-user limits.
5. Voice-driven on-site design. Walk the property, dictate observations, get real-time concepts.

12. Patent & IP Findings

Overall patent risk: Low-Moderate Freedom to operate: Clear (with caution areas noted)

Full analysis available in supporting/patent-landscape.md.

12.1 Patents by Feature Area

| Feature Area | Risk | Key Concern | Mitigation | |-------------|------|-------------|------------| | Conversational Image Editing | High | Adobe has 5+ granted US patents (2012-2024) including US11972757B2 | Differentiate via landscape domain + LLM function calling (not canonical intention mapping) | | AI Landscape Design Generation | High | Home Outside US12518067B2 covers AI landscape design with scoring | Use generative AI (not database-comparison scoring); architecturally distinct | | Text-Guided Diffusion Editing | Moderate | Google has 22+ patents on specific diffusion techniques | Use third-party model APIs (providers indemnify against technique-level claims) | | Credit-Based Pricing | Low | Standard business practice, no AI generation pricing patents | Well-established SaaS prior art | | Intent Routing (Multimodal) | Low | General assistant patents exist but not design-tool-scoped | Design-domain-specific classification is distinct | | Markup-Guided Generation | Low | No blocking patents in landscape context | OpenAI's annotation patent covers analysis, not generation guidance | | Follow-Up Clarification | Low | Closest patent (WO2024158398A1) has ceased | Consider defensive publication |

12.2 Key Patents Found

| # | Patent | Title | Assignee | Filed | Risk | Notes | |---|--------|-------|----------|-------|------|-------| | 1 | US11972757B2 | Conversational Image Editing & Enhancement | Adobe | 2018/2023 | High | Aesthetic scoring + canonical intention mapping in conversational editing | | 2 | US12518067B2 | System for Generating a Landscape Design | Home Outside | 2019 | High | AI landscape design with calculator + scoring engine — most direct competitive threat | | 3 | US10579737B2 | NL Image Editing Annotation Framework | Adobe | 2018 | High | NL-to-editing-command translation | | 4 | US11257491B2 | Voice Interaction for Image Editing | Adobe | 2018 | High | Voice-driven editing commands | | 5 | US20230230198A1 | Interactive Image Creation via NL Feedback (TiGAN) | Adobe | 2022 | Moderate | Multi-round conversational editing | | 6 | US20240037822A1 | Prompt-to-Prompt Editing with Cross-Attention | Google | 2022 | Moderate | Diffusion-based cross-attention manipulation | | 7 | US11983806B1 | Image Generation (inpainting/outpainting) | OpenAI | 2023 | Moderate | Core diffusion-based editing mechanics | | 8 | US12039431B1 | Multimodal ML Model Interaction (annotations) | OpenAI | 2023 | Moderate | GUI annotation-based region analysis | | 9 | EP4553759A2 | Multi-round Conversational Image Editing | Baidu | 2024 | Moderate | CN/EP jurisdiction — monitor for US continuation | | 10 | US9412366B2 | NL Image Spatial and Tonal Localization | Adobe | 2012 | Moderate | Foundational NL image editing (2012) |

12.3 Competitor IP Activity

Adobe: Very strong portfolio — 5+ granted US patents spanning 2012-2024 on conversational image editing. US11972757B2 (granted 2024) is the broadest, covering conversational editing with aesthetic scoring and canonical intention mapping. Strategy is aggressive and expanding. Primary patent concern.

Home Outside, Inc.: Single granted patent (US12518067B2, exp. 2041) specifically covering AI landscape design generation with calculator engine, scoring engine, and database comparison. Most direct competitive threat in the vertical, but describes a fundamentally different architecture (database-lookup scoring) than generative AI conversation.

Google/Alphabet: 22+ patents covering specific diffusion-based editing techniques (cross-attention manipulation, null-text inversion, hint-driven editing). Implementation-specific rather than application-layer — manageable through third-party API usage where providers handle licensing.

OpenAI: Three relevant patents covering inpainting mechanics, hierarchical text-to-image generation, and visual annotation interaction. Limited but growing portfolio.

Baidu / ByteDance: Recent filings covering multi-round conversational image editing (EP4553759A2, WO2025209146A1). Primarily CN/EP jurisdictions — monitor for US continuation filings.

Canva: No AI-specific patents found. Relies on third-party models and licensing.

LeanScaper / Stability AI / Midjourney: No relevant patents found.

12.4 Defensive Publications Found

TDCommons search returned 22 defensive publications (2022-2026) related to AI creative tools:

| # | Source | Title | Published | Overlap | |---|--------|-------|-----------|---------| | 1 | TDCommons | AI-driven Special Effects Generation Framework | 09/2025 | Partial — AI generation pipeline | | 2 | TDCommons | Assistive Interaction Mechanisms for AI-Powered Art | 09/2025 | Partial — accessible AI art interaction | | 3 | TDCommons | Conversational Agent for Physical Fulfillment of GenAI | 12/2025 | Partial — conversational agent → physical actions | | 4 | TDCommons | Location/Context-Specific Generative Multimedia | 08/2025 | Partial — property-specific generation | | 5 | TDCommons | AI-Based Creative Companion for Content Creation | 01/2022 | Low — general AI creative assistant prior art |

The defensive publication landscape is growing for AI creative tools but remains thin for landscape-design-specific AI. This is an opportunity for SimplyScapes to file.

12.5 Freedom-to-Operate Assessment

Favorable factors:

1. The full combination (conversational editing + domain object library + markup guidance + proactive clarification + credit billing + landscape-specific) is not claimed by any single patent.
2. SimplyScapes operates in landscape design (plant placement, hardscape, spatial layout) — fundamentally different from Adobe's photographic enhancement domain (exposure, contrast, color balance, aesthetic scores).
3. Credit/token billing, markup-guided generation, and intent routing for design tools remain largely unencumbered.
4. Key competitors Canva, LeanScaper, Midjourney have no relevant patents.

High-concern areas:

1. Adobe's conversational editing patents (US11972757B2 especially) — use LLM-based function calling for intent routing, NOT rule-based canonical intention mapping. Focus on landscape-specific intents rather than general photo editing. Avoid implementing aesthetic attribute scoring systems.
2. Home Outside's landscape design patent (US12518067B2) — use generative AI (diffusion models) for visual generation, NOT database-lookup design recommendation. Ensure the workflow is conversational/iterative, not score-and-improve. Do not implement a scoring engine comparing against online landscape databases.

Medium-concern areas: 3. Google's diffusion technique patents — use third-party model APIs where providers handle patent licensing. Avoid implementing specific patented techniques in custom code. 4. Baidu's multi-round conversational editing — primarily CN/EP jurisdiction. Monitor for US continuations.

Recommendations:

1. File defensive publications on TDCommons for: (a) property-context-aware landscape generation, (b) climate zone and plant hardiness integration in AI design, (c) multi-turn landscape refinement with spatial memory, (d) credit metering for domain-specific generation complexity, (e) object library integration with conversational editing.
2. Design around Adobe: use LLM function calling (not canonical intention mapping), focus on landscape domain (not photo editing), avoid aesthetic scoring systems.
3. Design around Home Outside: use generative AI conversation (not database-comparison scoring engines).
4. Establish quarterly patent watch for Adobe, Google, Baidu, ByteDance, and Home Outside in CPC G06T11/00, G06F30/13.
5. Flag US12518067B2 (Home Outside) and US11972757B2 (Adobe) for IP counsel review.

Note: Patent findings are NOT legal advice. Flag concerns for qualified counsel when warranted.

13. Academic & Open Source

Full analysis available in supporting/academic-open-source-scan.md.

13.1 Key Papers

| # | Type | Reference | Date | Relevance | |---|------|-----------|------|-----------| | 1 | Paper | DialogGen: Multi-modal Dialogue System for Multi-turn T2I | 2024 | Core architecture for MLLM-orchestrated multi-turn generation | | 2 | Paper | Talk2Image: Multi-Agent System for Multi-Turn Image Editing | 2025 | Multi-agent decomposition prevents intention drift | | 3 | Paper | TDRI: Two-Phase Dialogue Refinement for Interactive Generation | 2025 | Clarify-then-generate pattern for ambiguous requests | | 4 | Paper | Proactive Agents for Multi-Turn T2I Under Uncertainty | 2024 | Proactive clarifying questions improve satisfaction | | 5 | Paper | SmartEdit: Complex Instruction-based Image Editing (CVPR 2024) | 2024 | Joint MLLM+diffusion handles multi-step editing instructions | | 6 | Paper | BrushEdit: All-in-One Inpainting and Editing | 2024 | Free-form, multi-turn interactive editing with masks+text | | 7 | Paper | RouteLLM: Learning to Route LLMs with Preference Data | 2024 | 2x cost reduction routing between strong/weak models | | 8 | Paper | ToolACE: Winning the Points of LLM Function Calling (ICLR 2025) | 2024 | 8B models match GPT-4 on function calling with right training | | 9 | Benchmark | BFCL v3: Multi-turn Function Calling Benchmark | 2025 | JSON output reduces accuracy 27% vs. natural language reasoning | | 10 | Paper | Training-Free Sketch-Guided Diffusion | 2024 | Add sketch conditioning without retraining models | | 11 | Paper | AIdeation: Human-AI Collaborative Ideation (CHI 2025) | 2025 | Support both divergent and convergent design thinking | | 12 | Paper | Effects of Generative AI on Design Fixation (CHI 2024) | 2024 | Caution: AI generators increase fixation, reduce variety | | 13 | Paper | CVPR 2024 Instruction-guided Editing (Winning Solution) | 2024 | Pipeline: classify → identify region → mask → inpaint | | 14 | Survey | Multi-modal Intent Recognition Survey (EMNLP 2025) | 2025 | Single-modal data insufficient for complex intent classification | | 15 | Paper | Improving LLM Function Calling via Structured Templates (EMNLP 2025) | 2025 | Reasoning chain before tool calls improves accuracy |

13.2 Open Source Projects

| # | Project | Description | Relevance | |---|---------|-------------|-----------| | 1 | RouteLLM | Dynamic routing between strong/weak LLMs | Cost optimization for intent routing | | 2 | BrushNet | Plug-and-play inpainting for diffusion models | Mask-based editing without retraining | | 3 | Gorilla | LLM for API/function calling | Retriever Aware Training for changing tool sets | | 4 | Open Pencil | AI-native design editor (87 tools) | Reference for AI tool architecture in design | | 5 | Open Canvas | LangChain collaborative writing/coding agents | Content + reflection agent architecture | | 6 | Vercel AI SDK 6 | Provider-agnostic AI toolkit for Next.js | Direct integration path — tool calling, streaming, agents | | 7 | LangGraph | Stateful agent graphs with human-in-the-loop | State machine pattern for chat flow orchestration |

13.3 Key Architectural Patterns from Literature

1. Hybrid Intent Router (RouteLLM + EMNLP 2025 survey) Lightweight classifier as first stage → route to specialized agents. 2x cost reduction without quality loss. Directly maps to SimplyScapes' two-phase architecture.

2. Multi-Agent Decomposition (Talk2Image) Separate intention parser, task decomposer, specialized executors, and evaluator agents. Prevents intention drift in multi-turn conversations — critical for iterative design editing.

3. Clarify-Before-Generate (TDRI + Proactive Agents) When intent is ambiguous, ask targeted clarifying questions rather than guessing. Improved user satisfaction documented. Essential for landscape design where "make it look better" needs disambiguation.

4. Pipeline Editing (CVPR 2024 winning solution) Classify edit type → identify target region → generate mask → inpaint. More reliable than end-to-end approaches. Maps to SimplyScapes' mode-based architecture.

5. Structured Reasoning Before Tool Calls (EMNLP 2025) Generate reasoning chain before structured output. JSON output during reasoning reduces accuracy by 27%. Use natural language intermediate reasoning, then emit structured tool calls.

6. Anti-Fixation Design (CHI 2024) AI generators increase design fixation — users produce fewer ideas with less variety. Counter by proactively offering alternatives and variations, not just refining the first generated image.

7. Training-Free Sketch Conditioning (arXiv 2024) Add sketch/markup guidance to existing diffusion models without retraining. Latent optimization at each denoising step ensures adherence to spatial structure. Directly enables markup-guided generation.

13.4 Pricing Model Research

Academic and industry research confirms credit-based pricing is the dominant emerging model:

• Credit models grew 126% YoY (35 to 79 companies in PricingSaaS 500 Index)
• Seat-based pricing dropped from 21% to 15%
• Top blocker: Customer anxiety about unpredictable credit burn rates
• Mitigation: Display credit cost before each action, provide usage dashboards
• Evolution path: Cost-plus credits → value-aligned credits → outcome-based pricing (McKinsey)

---

Part IV: Synthesis

14. Opportunity Map

Validated Patterns (safe to build on)

These approaches are used by multiple products with no patent barriers. They represent table-stakes expectations:

1. Credit-based AI billing. Used by Canva, Figma, Midjourney, Adobe, Runway, ChatGPT. No IP barriers. Standard SaaS practice. SimplyScapes' 10-credit-per-image model aligns with market norms.

2. Mode-based AI interfaces. Canva, Figma, and Adobe all organize AI capabilities into distinct modes/tools (generate, edit, erase, expand). The mode dropdown pattern is validated.

3. Mask-based inpainting. Every design tool with AI editing uses user-defined masks to constrain generation. Well-established technique with open-source implementations (BrushNet, BrushEdit).

4. Conversation history for context. ChatGPT, Midjourney, and Canva all maintain multi-turn context. Sliding window approaches are standard.

5. Provider abstraction. Vercel AI SDK, LangChain, and every production AI app abstract the model provider. The AIProvider interface pattern is universal.

Differentiation Opportunities (where to innovate)

These combine market gaps, transferable patterns, and patent-free zones into genuine competitive advantages:

1. Conversational design partner with domain object library

   • Market gap: No landscape AI product offers conversational flow that routes between generation, clarification, and structured object selection from a curated catalog.
   • Inspiration: ChatGPT's multi-turn editing + Figma's structured output + Canva's contextual placement.
   • Patent risk: Low. The combination of conversational editing with domain-specific object library is unclaimed.
   • Why it's different: Every competitor generates flat images. SimplyScapes can generate images AND place structured design objects (plants with metadata, hardscape with pricing) — bridging the gap between inspiration and actionable design.

2. Two-phase intent routing with function calling

   • Market gap: No landscape tool classifies user intent before routing to specialized agents. All competitors are single-path (prompt → image).
   • Inspiration: RouteLLM (2x cost reduction), EMNLP 2025 structured reasoning, Gemini function calling.
   • Patent risk: Low. Intent routing patents are scoped to general voice assistants, not design tools.
   • Why it's different: A cheap text model decides what the user wants (10x cheaper than always running image generation). The mode dropdown provides a hint but doesn't constrain — the AI can override when intent doesn't match mode.

3. Proactive clarification with visual choices

   • Market gap: No product in the vertical proactively asks follow-up questions with thumbnail-rich choices from a product catalog.
   • Inspiration: Proactive Agents paper (2024), TDRI two-phase dialogue (2025), ChatGPT's conversational refinement.
   • Patent risk: Low. Closest patent (WO2024158398A1) has ceased.
   • Why it's different: Instead of guessing what "a water feature" means, the system asks "which style?" and presents 6 fountains from the object library with thumbnails and prices. The user picks, and the selection (including metadata) feeds back to Gemini for contextually accurate generation.

4. Markup-as-generation-guidance

   • Market gap: No landscape tool lets users draw on the canvas to guide AI placement. Competitors use only text prompts or predefined masks.
   • Inspiration: Training-free sketch-guided diffusion (2024), CVPR 2024 pipeline approach.
   • Patent risk: Low-Moderate. OpenAI's annotation patent covers analysis, not generation guidance. Differentiate clearly.
   • Why it's different: Users draw rough shapes and arrows on the design, and these spatial cues guide where and what Gemini generates. Landscapers think spatially — this matches their workflow.

5. $995 gap between free AI and human design

   • Market gap: iScape offers free AI inspiration images. Yardzen charges $999+ for human-designed plans. Nothing exists between.
   • Inspiration: Canva's AI-assisted DIY model (user + AI together for $15/mo).
   • Patent risk: None — this is market positioning.
   • Why it's different: SimplyScapes can offer AI-assisted design at $29-79/mo that produces structured, actionable landscape plans — not just pretty pictures. The object library, plant database, and structured output bridge the gap between "inspiration" and "professional design."

Caution Zones (promising but constrained)

1. Adobe's conversational editing portfolio — Adobe holds 5+ granted US patents covering conversational image editing (US11972757B2 is broadest — aesthetic scoring, canonical intention mapping, iterative suggestion). Do not implement: rule-based canonical intention mapping, aesthetic attribute scoring systems, or NL-to-editing-command translation. Alternative: Use LLM function calling for intent routing (architecturally distinct from Adobe's approach). Focus on landscape-specific intents (plant placement, hardscape, seasonal) rather than photo editing (exposure, contrast, color).

2. Home Outside's landscape design patent — US12518067B2 (granted, exp. 2041) covers AI landscape design with calculator engine, scoring engine, and database comparison. Do not implement: a scoring engine that retrieves landscape data from online databases and compares against existing designs. Alternative: Use generative AI (diffusion models via Gemini API) for visual generation and conversational refinement. The approach is architecturally distinct — conversation-driven generation vs. database-comparison scoring.

3. Autonomous multi-step editing — The Talk2Image multi-agent pattern (parser → generator → evaluator) is academically validated but production-complex. Multi-agent systems add latency and failure modes. Alternative: Start with the two-phase router (classify → execute) and only decompose into more agents if single-model quality degrades on complex requests.

4. Sketch-to-design generation — Training-free sketch conditioning is research-validated but depends on Stable Diffusion / FLUX architectures. Gemini's image generation pipeline may not support external conditioning inputs. Alternative: Treat markup as metadata that enriches the text prompt ("generate a fountain in the upper-right quadrant, approximately 3 feet wide") rather than as a spatial conditioning signal.

15. The Technical Landscape

The architecture that emerges from research is a two-phase, provider-abstracted pipeline with these key components:

User Input (text + image + mask + markup + selection)
  │
  ├─ Phase 1: Intent Classification (Gemini 2.5 Flash — text only)
  │    ├─ Function calling with 6 tool declarations
  │    ├─ Mode hint from dropdown informs but doesn't constrain
  │    ├─ Result: generate_image | search_objects | answer_question |
  │    │          ask_followup | request_markup | erase_region
  │    └─ Cost: ~$0.001 per classification
  │
  ├─ Phase 2: Execution (model depends on intent)
  │    ├─ generate_image → Gemini 2.5 Flash Image ($0.039/image)
  │    ├─ search_objects → Hasura query (free)
  │    ├─ answer_question → Gemini 2.5 Flash text ($0.001)
  │    ├─ ask_followup → Return choices JSON (free)
  │    ├─ request_markup → Return instruction (free)
  │    └─ erase_region → Gemini 2.5 Flash Image ($0.039)
  │
  ├─ Credit Deduction (atomic Hasura operation)
  │    ├─ Pre-deduct before generation
  │    ├─ Refund on failure
  │    └─ 10 credits (image) / 2 credits (text) / 0 (follow-up)
  │
  └─ Conversation Persistence (Hasura)
       ├─ ai_conversation + ai_conversation_turn tables
       ├─ 5-8 turn sliding window for context replay
       └─ Analytics events for usage tracking

Key technical decisions validated by research:

1. Synchronous-first, streaming-ready. Image generation is inherently synchronous (Gemini returns the full image). Text responses could stream but the added complexity isn't justified for short answers. Build the API synchronous with a stream: boolean parameter for future use.

2. App Router migration. New endpoint at src/app/api/v1/ai/generate/route.ts. Keep the existing Pages Router endpoint (src/pages/api/ai-gen.ts) for v1 compatibility. Both can coexist in Next.js.

3. Gemini-first, provider-agnostic. The AIProvider interface abstracts Gemini specifics. Start with Gemini (best price/quality for image generation as of March 2026). The interface allows swapping in other providers or routing between them based on task.

4. Function calling for intent, not for generation. Gemini's function calling and image generation capabilities cannot be combined in a single API call. The two-phase architecture turns this limitation into a strength: cheap intent classification, expensive generation only when needed.

5. Hasura for state, not for logic. Conversation persistence, credit management, and analytics all go through Hasura. Business logic (intent routing, prompt construction, provider selection) stays in the Next.js API layer.

Migration requirement: Gemini 2.0 Flash shuts down June 1, 2026. Migrate to Gemini 2.5 Flash before then. The 2.5 Flash model is cheaper ($0.30 vs. $0.15/1M input tokens) and supports both text and image generation with a single model family.

16. Open Questions

• [ ] Gemini 2.5 Flash Image quality vs. 3.1 Flash Image. The 3.1 model costs 72% more ($0.067 vs. $0.039/image). Is the quality difference worth it for landscape-specific generation? Needs A/B testing with real design prompts.
• [ ] Context window for image history. Gemini's context window accepts multiple images, but each 1024px image costs ~258 input tokens. How many reference images can be included before quality degrades or costs spike? Needs empirical testing.
• [ ] Object library search quality. The search_objects function call depends on Hasura's text search or pg_trgm similarity. Is this sufficient for natural language queries like "something with purple flowers that grows in shade"? May need vector search (pgvector) or Gemini-mediated search.
• [ ] Markup interpretation fidelity. When users draw on the canvas, how reliably can Gemini interpret spatial annotations in a reference image? Early testing needed with annotated design images.
• [ ] Whitelabel credit isolation. The instance_id field in the API supports multi-tenant credit pools, but the Hasura permission model needs validation. Can workspace admins see credit usage across instances?
• [ ] Rate limiting granularity. 30 req/min per workspace is proposed, but image generation requests are expensive and slow (~3-5s). Should image-generation requests have a separate, lower limit (e.g., 10/min)?
• [ ] Conversation cleanup policy. How long should conversations persist? Storage is cheap, but old conversation context could confuse analytics. Consider auto-archiving after 30 days of inactivity.
• [ ] Anti-fixation UX. CHI 2024 research shows AI generators increase design fixation. How should the UI counteract this — offer unprompted variations? Show "try something different" affordances?

17. Opportunity Assessment

Novelty: High. No product in the landscape vertical offers conversational AI design with structured object library integration, intent-based routing, proactive clarification, and markup-guided generation. The closest analog is ChatGPT's image editing, but it produces flat images with no domain knowledge, structured data, or object placement. The combination of capabilities is novel and defensible.

Feasibility: High. All core components are buildable with current technology:

• Gemini 2.5 Flash handles both intent classification (function calling) and image generation
• The two-phase architecture is well-documented in academic literature (RouteLLM, EMNLP 2025)
• Credit billing is standard SaaS (Stripe native credits)
• Hasura handles conversation persistence and credit state
• The existing v1 inpainting pipeline proves the core image generation loop works

The main technical risk is Gemini's image generation quality for landscape-specific scenes. Early testing with real design prompts is the fastest way to validate.

Impact: Very High. This transforms SimplyScapes from "a design tool with an AI feature" to "an AI-powered design partner." The $995 gap between free AI inspiration (iScape) and human design services (Yardzen) is a massive, unserved market. A conversational AI design partner at $29-79/mo could capture landscape professionals who want AI assistance but need structured, actionable output — not just pretty pictures.

Timeline:

• Phase 1 (4-6 weeks): Two-phase intent routing, basic credit system, conversation persistence. Extends v1 from single-turn to multi-turn with intent classification.
• Phase 2 (4-6 weeks): Object library integration with search_objects and ask_followup response types. Proactive clarification with visual choices.
• Phase 3 (4-6 weeks): Markup-guided generation, provider abstraction layer, advanced credit management (workspace pools, credit packs, Stripe integration).
• Ongoing: System prompt refinement, model quality testing, analytics-driven optimization.

---

Part V: UX Direction Pivot — Visual Designer AI Tab Workflow

Pivot date: 2026-05-01. Driver: brief.md (Visual Designer AI Tab — Guided Workflow). This part supersedes the chat-API direction in Parts I–IV for UX purposes while preserving Parts I–IV as backend infrastructure reference. The prior research (intent routing, function calling, credit pricing, patent landscape) remains valid and useful — it describes the machinery underneath. This part addresses the interaction layer the user actually sees: a guided, click-driven, design-system-first workflow inspired by Anthropic's Claude Design and the canonical professional landscape design process.

Note on numbering. The existing "Part IV: Synthesis" above is the synthesis of the chat-API direction. This new section is numbered Part V to preserve the integrity of the prior work, even though it is functionally the new "Part IV" by topic. The brief's research-questions framing maps to Sections A–E below.

A. Professional Designer Workflow

Full treatment in research/supporting/landscape-designer-workflow.md. Highlights below.

The American Society of Landscape Architects (ASLA Documentation Standards) codifies five canonical phases that landscape designers and architects use across residential and commercial projects. Every accredited landscape architecture program teaches them; every professional firm follows them; university extension programs (NC State, University of Florida) teach the homeowner version of the same process.

| # | Phase | Primary Question Answered | AI Tab Mapping | |---|-----------------------------|-----------------------------------------------|---------------------------------------------| | 1 | Site Analysis / Pre-Design | "What is here, and what does it constrain?" | Out of scope (other features handle this) | | 2 | Programming | "What functions must the design serve?" | Lead-in questionnaire | | 3 | Schematic Design | "What is the spatial organization?" | Steps 1–2 (Prepare site, Define spaces) | | 4 | Design Development | "What does it actually look like?" | Steps 3–4 (Plantings, Decor) | | 5 | Construction Documentation | "How does it get built?" | Out of scope (Site Planner, Document Gen) |

Three takeaways:

1. The brief's five-step flow is professionally correct. Prepare → Define spaces → Add base plantings → Decor + finishing → Views maps cleanly to programming/schematic/early-DD phases. We don't have to invent a workflow — we follow the one designers already use.

2. The "bones to jewelry" sequencing rule. Designers go top-down by size and permanence: trees (bones) → foundation shrubs (structure) → perennials (body) → lighting and decor (jewelry). Amateurs do the opposite, which is why amateur designs feel fragmented. The brief's step order matches the professional sequence — make sure UX preserves that ordering.

3. Multiple concepts is the schematic-phase standard. Professional designers always present 2–3 alternatives. Decorilla and Modsy operationalize this as "two concept boards." The existing version-history UI is the SimplyScapes equivalent. Surface it as a first-class concept-comparison surface, not as an undo affordance.

B. Adjacent AI Design Tools — Onboarding Patterns

Full treatment in research/supporting/ai-design-tool-onboarding.md. Deep dive on Anthropic Claude Design + comparative scan of nine other tools.

B.1 Claude Design — the closest analogue

Anthropic's Claude Design is the only major AI design tool that puts a design-system-first onboarding ahead of any creative work. The flow has three stages:

1. Organization Onboarding (one-time). User uploads brand assets — GitHub links, Figma files, font folders, logos, slide decks, screenshots, free-form style notes. Claude reads them and extracts a four-part design system: color palette, typography, components, layout patterns. User reviews, edits if needed, and clicks Publish. Every project after that inherits the system automatically.

2. Project Creation. User clicks "New project," picks a fidelity level, types a prompt covering goal/layout/content/audience, optionally attaches references. Claude may pause and ask clarifying questions when the prompt is too thin — but only conditionally. Power users with detailed prompts skip the questioning.

3. Iteration. Two-pane UI (chat panel left, canvas right) with four parallel refinement channels: chat for broad changes, inline comments on canvas elements, direct text editing, and contextual sliders for spacing/color/layout.

Reported outcomes: Brilliant went from 20+ prompts to 2 prompts; Datadog ships prototypes "before anyone leaves the room"; what used to take a week now happens in single conversations. The speed advantage is the design system inheritance — every output is on-brand without manual brand work. The same multiplier applies to landscape: a "Cottage" pick delivers coherent vocabulary across plants, materials, color, and density without specifying each.

Key UX choices to borrow:

• Asset upload over text questions — when artifacts exist, infer; only ask questions when nothing else suffices
• Conditional clarification — pause for questions only when input is thin
• Parallel refinement channels — chat + inline + direct + sliders, each for different change types
• Named system inheritance — pick once, propagates to every step

B.2 Comparative scan — the other nine tools

| Tool | Onboarding pattern | Transferable | |----------------------|---------------------------------------------|---------------------------------------| | Figma Make | Inherits Figma library; no separate setup | Multi-modal input (text + image + capture) | | Spacely AI | Preset Mode vs. Prompt Mode toggle | Adopt for paid-tier prompt access | | Modsy (defunct) | 5-min image-pick style quiz → algorithm | Photo picks beat text questions | | Havenly / Decorilla | Style quiz → designer match | Named style outputs ("Cottage") | | Yardzen | 9-step intake (~3 hrs total) | Multi-step works; 3 hrs is too long | | Neighborbrite | Pick from 16 styles, generate (single shot) | Fast lane is expected; gap = workflow | | Higharc | AI baseline + buyer configurator on top | AI baseline + drag/drop layer | | Maket | Specify constraints → multiple variations | Multiple variations is now expected |

B.3 The onboarding spectrum

Light (no setup)                                              Heavy (full system)
|------------------|------------------|------------------|------------------|
                   |                  |                  |
       Neighborbrite          Modsy / Havenly           Claude Design
       Garden AI              Yardzen
       (single shot)          (5–60 min quiz)           (full org-level
                                                         system, asset
                                                         extraction)

Where SimplyScapes should sit. Between Modsy/Havenly and Claude Design. More structured than Neighborbrite (because pros need a real workflow, not a one-shot transform), less heavy than Yardzen (because we need <5-minute time-to-first-render). The pre-baked design system picker is the SimplyScapes innovation that fills this gap.

C. Landscape Design Systems

Full treatment in research/supporting/landscape-design-systems.md.

C.1 The eight-attribute model

Translating the web/UI design-system concept (color, typography, spacing, components) to landscape design produces eight attribute categories:

| # | Attribute | Examples of Values | |---|------------------------|----------------------------------------------------------| | 1 | Plant palette | Core species (5–10), accents (3–5), filler (2–4), family bias | | 2 | Hardscape vocabulary | Primary stone, secondary material, edge treatment, vertical structures | | 3 | Formality | Formal / Semi-formal / Transitional / Informal / Naturalistic | | 4 | Density | Minimal / Moderate / Lush | | 5 | Color & seasonality | Multi-colored / Whites / Earth tones / Cool grays + greens; spring-dominant / multi-season / evergreen | | 6 | Maintenance load | Very low (≤2 hrs/mo) / Low / Moderate / High / Very high (>20 hrs/mo) | | 7 | Water use | Xeric / Low / Moderate / High | | 8 | Form language | Geometric / Curvilinear / Architectural / Organic / Cultural-specific |

Sources: University of Florida MG086 on plant palette sizes, Bullard Bollards on hardscape material limits, Yardzen style guide for cross-style attribute data.

The user does not need to understand this taxonomy. They pick one named system and the eight attributes are pre-set. The taxonomy is the schema underneath — the production prompt fragment.

C.2 The twelve recommended pre-baked systems

We propose shipping 12 systems organized into three groups of four (4×3 grid). Twelve is deliberately less than Yardzen's 21 — past 8–10 choices, users feel paralyzed rather than empowered, and 12 well-tuned systems produce better outputs than 21 rough ones.

Group A: The Essentials (universally recognized, high pick rate)

1. Cottage — lush, romantic, multi-colored; roses, lavender, perennials packed shoulder-to-shoulder
2. Modern — clean, calm, intentional; architectural plants, large pavers, negative space
3. Naturalistic — looks wild but is engineered; native masses, pollinator-rich, low maintenance
4. Formal Traditional — geometric, symmetrical, classic; clipped hedges, axial paths, boxwood/yew/hydrangea

Group B: Regional / Climate-Driven (climate-anchored defaults, available everywhere) 5. Xeriscape — drought-proof; decomposed granite, agave, lavender, very low water 6. Mediterranean — lavender, olive, rosemary; stone walls, gravel courtyards 7. Tropical — bold, escapist; large-leafed plants, vivid color, dense layering 8. Coastal — salt-tolerant, breezy, weathered; grasses, lavender, weathered wood

Group C: Specialty / Cultural / Lifestyle (recognizable, named, meaningful) 9. Japanese Zen — contemplative, restrained; mossy ground, raked gravel, Japanese maples 10. Modern Farmhouse — modern's quieter cousin; grasses, herringbone brick, split-rail fencing 11. Prairie / Native — habitat-first; tall grasses and wildflowers, almost no hardscape 12. Rustic — warm, textural, organic; wood, mulch, gravel, full plantings

Each system's attribute matrix and starter prompt fragment is in landscape-design-systems.md § 4.

C.3 The system → prompt fragment

When a user picks "Cottage," the AI tab assembles a structured prompt fragment that becomes the constant context for every subsequent step:

Style: English Cottage Garden
Plant palette: roses, lavender, salvia, foxglove, hydrangea, climbing
vines; emphasis on multi-colored perennials in dense plantings.
Hardscape: curved brick paths, weathered wood structures,
picket-style fencing.
Color palette: bright multi-colored with whites and pinks dominant;
warm wood tones for hardscape.
Formality: informal — curving bed shapes, organic massing,
no straight lines.
Density: lush — beds packed with overlapping plantings.
Maintenance: high — mature, established garden look.
Form language: curvilinear — flowing curves, soft edges.

This fragment concatenates with the per-step intent ("add base plantings," "remove that overgrown shrub") to produce the final generation prompt. The design system is the constant context; the step intent is the variable.

D. Sample Questionnaire & Flow Content

Material for the UX Designer to refine. This is research scaffolding, not polished UX copy.

D.1 Lead-in questionnaire (Programming-phase content)

The programming phase in professional landscape design captures who-uses-the-space, what-they-do-there, and what-constraints-apply. The mobile-friendly version is 5–8 lightweight click steps (modeled on Yardzen's intake compressed by 3–5x). All questions support image-pick answers where viable; text fallback for the few that don't.

Q1. Who uses this space? (multi-select, image-driven where possible)

• Just me / partner
• Kids
• Pets
• Older parents / accessibility considerations
• Frequent guests / entertaining

Q2. What do you want to do here? (multi-select)

• Relax / read / quiet space
• Entertain / gather around a fire pit or table
• Cook outside (BBQ / outdoor kitchen)
• Garden — flowers
• Garden — vegetables / herbs
• Play (kids / pets)
• Privacy from neighbors / road
• Something to look at from indoors

Q3. How much yard work do you actually want to do? (single-select, slider)

• Very little — set it and forget it
• Some — weekends are fine
• A lot — gardening is a hobby
• I have a landscaping company

Q4. Water situation. (single-select)

• Drought / very limited irrigation
• Some restrictions — drip system OK
• No restrictions — sprinklers, regular watering OK

Q5. What's your home's vibe? (single-select, image-pick — show 4–6 home photos)

• Modern / clean lines
• Traditional / classic
• Farmhouse / cottage
• Rustic / craftsman
• Mediterranean / Spanish

Q6. (Optional) What do you want to keep out? (multi-select)

• Pet-toxic plants
• Plants with thorns
• Plants my partner is allergic to (specify)
• (None — surprise me)

Q7. (Optional, free text) "Anything else I should know about how you use this space?"

The flow ends with the design-system picker (4×3 grid of named systems with hero images), with the user's questionnaire answers used to soft-rank the systems — the best-fit ones float to the top of the grid, the rest stay available below.

D.2 Step-by-step flow copy (sketches — not final)

Sketches of microcopy at each of the brief's five steps, calibrated for non-designers:

Step 1 — Prepare the site.

"Let's clear the canvas. What's in your way?"

Click an object on your photo to remove it (or pick from below).

[ Overgrown shrubs ] [ Old trees ] [ Vehicles ] [ Junk / clutter ] [ All plants — start fresh ] [ Concrete / hardscape ]

Step 2 — Define spaces and layout.

"Now let's shape your spaces."

[ Reduce lawn to a center circle ] [ Cut a deep planting bed ] [ Add a patio area ] [ Add a path ] [ Build a privacy screen ] [ Skip — keep current shape ]

Step 3 — Add base plantings.

"Time to plant. Drag from the panel, or pick a layer to fill."

Layer: [ Trees ] [ Foundation shrubs ] [ Perennials ] [ Groundcover ]

Or: [ Auto-fill this bed with Cottage plantings ]

[ Apply staged plants — flatten ]

Step 4 — Decor + finishing touches.

"Add the jewelry."

[ Path lighting ] [ Up-lighting on trees ] [ Fountain / water feature ] [ Fire pit ] [ Pergola / arbor ] [ Garden art ]

Step 5 — Views.

"See it from every angle."

[ Top view ] [ Side view ] [ Front entry ] [ From the kitchen window ]

[ Toggle plant labels ]

Tone reference: Yardzen's quizzes ("These are fun little quizzes and are really helpful for someone like me who has no idea what I really want or need" — Lela Burris review). Conversational but action-oriented. Fewer than 12 words per button.

D.3 Design system picker copy

Each of the 12 systems gets a one-line description (a "tagline") in the picker grid. Examples:

| System | Tagline | |---------------------|---------------------------------------------------------------| | Cottage | Lush, romantic, multi-colored. Packed beds, curving paths. | | Modern | Clean, calm, intentional. Architectural plants and big pavers. | | Naturalistic | Looks wild but engineered. Native plants, pollinator-rich. | | Formal Traditional | Geometric and classic. Clipped hedges, symmetrical beds. | | Xeriscape | Drought-proof. Climate-adapted, very little water. | | Mediterranean | Lavender, olive, gravel courtyards. Warm and timeless. | | Tropical | Bold and escapist. Big leaves, bright color. | | Coastal | Breezy and weathered. Grasses and driftwood tones. | | Japanese Zen | Contemplative. Moss, gravel, Japanese maples. | | Modern Farmhouse | Modern's quieter cousin. Grasses, brick, split rail. | | Prairie / Native | Habitat-first. Tall grasses and wildflowers, no lawn. | | Rustic | Warm and textural. Wood, mulch, full plantings. |

E. Synthesis — Recommended Workflow Direction

A high-level recommendation. The full UX spec — wireframes, interaction patterns, screen-by-screen copy, Claude Design composite prompt — is the next phase deliverable (the UX Designer's job, not this research round's).

The recommended workflow shape:

The Visual Designer AI tab opens with a design-system picker first, the way Claude Design opens with the design system. The picker is a 4×3 grid of named systems (Cottage, Modern, etc.) with hero images. Below it, two affordances: (a) "Use my last design system" for returning users, and (b) "Take a 60-second style quiz to find your fit" for the indecisive — the quiz being the lead-in questionnaire from Section D.1.

Once a system is chosen, the user is in the multi-step guided flow: Prepare site → Define spaces → Add base plantings → Decor → Views. Each step is skippable and reorderable (the brief's non-linearity requirement). Each step exposes the same set of refinement channels Claude Design uses — chat for broad changes, inline element controls for local edits, and the existing object library for drag-and-drop.

The design system functions as constant context through every step — the prompt fragment from Section C.3 prepends every generation, ensuring "Cottage" coherence whether the user is removing a shrub, placing a fountain, or choosing a path material. The user never has to retype it.

The version history UI is elevated from undo affordance to concept comparison surface — at any point the user can spawn a parallel concept by branching the version tree, develop both in parallel, and compare side-by-side. This implements the professional designer's "present 2–3 concepts" pattern as an emergent property of the existing UI rather than as a new bolt-on.

Free-text prompt entry is available at every step but paid-tier-gated and not the primary path — Spacely AI's Preset/Prompt mode toggle is the model. The default path is click-driven; the prompt path is a power user's escape hatch.

Why this works:

1. Time-to-first-render under 60 seconds. Pick a system → see a Cottage-flavored render of your yard. Faster than Neighborbrite's single-shot model and dramatically faster than Yardzen's 3-hour intake.

2. Coherent outputs by default. Every step's generation inherits the design-system context. No drift between steps. No Tropical plants appearing in a Modern composition.

3. Professional structure, novice surface. The five-step flow is the canonical landscape design process compressed into a homeowner-friendly UI. Pros recognize the workflow; novices don't have to.

4. Pro escape hatches preserved. Power users get prompt access (paid), drag-and-drop from the object library, and direct version-tree branching for parallel concepts.

5. Defensible against vertical competitors. Neighborbrite and Garden AI are single-shot. Yardzen is human-driven. Nobody else combines design-system-first + multi-step + drag-and-drop + version branching in landscaping.

Open questions for the UX phase

These are deliberately not answered here — the UX Designer should resolve them in collaboration with the Product Lead:

1. Picker-first vs. quiz-first? Should the default landing be the 4×3 grid of named systems, or the 60-second style quiz that recommends one? Either is defensible. Need user testing.

2. System mismatch handling. When a user picks "Tropical" but lives in Zone 4, what happens? Soft-flag with a "consider X instead" suggestion? Hard-block? Quietly substitute hardy alternatives in the prompt? Recommend soft-flag for v1.

3. Step skipping affordances. The brief requires non-linear flow. How explicit should "skip this step" be? A skip button on every step? An overall stepper UI showing all five with checkmarks? Need wireframes to test.

4. Drag-and-drop staging visibility. Staged objects flatten on "Apply." How visible is the staging layer to the user? Outlined ghosts? Half-opacity? An explicit "X items staged — apply or cancel" affordance? UX exploration needed.

5. Concept branching surface. Does the user explicitly fork a concept ("save this and try another"), or does the version history accumulate every generation and allow reverting/branching after the fact? Two different mental models.

6. Style quiz exit behavior. When the quiz ends, does it auto-pick the recommended system, or does it land on the picker grid with the recommendation highlighted? The latter preserves user agency.

7. The "Views" step. Camera angle changes are technically nontrivial (single-image-to-multi-view is a hard generative AI problem). What's actually buildable in v1? Plant labels overlay is straightforward; novel camera angles may need to be a "preview/teaser" feature behind a coming-soon flag.

8. Returning user flow. A user who has used the AI tab before — do they land on their last design system, or get the picker again? Probably the last system, with an obvious "change system" affordance. Needs testing.

These questions are the UX phase's job. This research round delivers the what (workflow shape, design systems, attributes, sample copy); the UX Designer delivers the how (wireframes, screen-by-screen interaction patterns, the Claude Design composite prompt for prototyping).

---

Part VI: Delta — Auto-Classification, Property Profile, Multi-Image

Pivot date: 2026-05-01. Driver: brief.md revisions adding panel-only constraint, auto-classification (home architectural style + existing landscape style), Property Design Profile entity, and multi-image reference subtypes.

Parts I–V remain valid. This is a delta round answering Research Questions 5–8 from the revised brief, focused on three problems the earlier rounds did not need to solve: (a) auto-classifying home and landscape style at site-analysis time, (b) inheriting a design profile across multiple visual designs of the same property, and (c) handling multi-image references in the panel.

Two supporting documents go deep: home-style-classification.md covers Sections A and B; multi-image-design-patterns.md covers Section C.

Section A: Home Architectural Style Classification

The decision. Use a few-shot Gemini 2.5 Pro Vision prompt for home architectural style classification, not a custom CNN. Run it as part of the existing site-analysis call — no new system dependency, no separate model-ops surface. Full reasoning in home-style-classification.md § A.

The benchmarks. Home-style classification has decades of academic prior art. The standard dataset is Architectural Style (Xu/ECCV 2014) — ~4,800 images, 25 world-architecture classes, 46% top-1 accuracy on the 2014 baseline and roughly 70-80% on modern attention-CNN approaches per Springer 2022. The directly relevant US-residential prior art is McIntire's home-type project, which fine-tuned ResNet50 on ~100 Bing-scraped images per class across 8 styles (Cape Cod, Colonial, Victorian, Tudor, Craftsman, Spanish, Mid-Century, plus a dropped Edwardian) and reached 76% accuracy. The Stanford CS231n project Classifying U.S. Houses sits at the same baseline. A dedicated CNN trained on a properly curated US-residential dataset (10K+ images, 12 classes) could plausibly reach 75–85% top-1, but the data-collection cost is real and the result is a static model that ages out as fashions shift.

Why a vision LLM wins here. Three reasons make the few-shot vision-LLM path strictly dominate the custom-CNN path for SimplyScapes:

1. Accuracy is in the same range. Frontier vision LLMs hit 70-85% accuracy on fine-grained zero-shot / few-shot classification tasks when given 3-5 exemplars per class. Cross-domain evidence: the LM Council May 2026 vision benchmarks put Gemini 2.5 Pro at 79.6% on MMMU; the brain MRI sequence classification benchmark shows GPT-4o at 97.7% and Gemini 2.5 Pro at 93.1% on a domain- specific 10-class problem with no domain training. Architectural style is messier than MRI but the same order of magnitude. Per arXiv 2510.03903 on zero-shot fine-grained classification, modern LVLM approaches close most of the gap with task-specific models.
2. The site-analysis pipeline already runs Gemini Vision. Adding a home-style field is one more entry in the structured response. No new infrastructure.
3. It evolves with model upgrades for free. Gemini 2.5 → 3.0 is a quality boost with no retraining cost.

The label set. 12 US-residential classes, each mapped to 1-3 preferred design systems for the "Recommended for your home" ribbon:

| Home class | Recommended systems | |-----------|---------------------| | Craftsman / Bungalow | Cottage, Naturalistic, Rustic | | Ranch | Modern Farmhouse, Xeriscape, Modern | | Modern / Contemporary | Modern, Japanese Zen, Xeriscape | | Colonial / Traditional | Formal Traditional, Cottage | | Tudor | Formal Traditional, Rustic | | Mediterranean / Spanish | Mediterranean, Xeriscape | | Mid-Century Modern | Modern, Japanese Zen | | Farmhouse / Modern Farmhouse | Modern Farmhouse, Cottage, Prairie/Native | | Cape Cod | Cottage, Coastal | | Victorian | Cottage, Formal Traditional | | Coastal / Beach | Coastal, Tropical | | Other / Unrecognized | (no recommendation) |

Full detail in home-style-classification.md § A.3.

Confidence-thresholding pattern. Three thresholds drive UI behavior:

| Confidence | UI behavior | |-----------|-------------| | High (>0.80) | Silent passive label on the picker, no confirm step | | Medium (0.50–0.80) | Soft prompt with chip-menu correction to top-3 alternates | | Low (<0.50) | Skip the recommendation entirely |

The threshold values are starting estimates. Calibration must happen on a labeled validation set of ~200 real Visual Designer photos before launch. The "Confidence UI" pattern recommends bucketing into "looks like" / "could be" / "unsure" affordances rather than showing raw percentages — which holds for SimplyScapes too. Detail in home-style-classification.md § A.4.

The yard-photo edge case. Real Visual Designer photos are not front-elevation architectural shots. They're taken from the yard at oblique angles, with the home occupying 20-50% of the frame and often truncated by trees. This breaks the academic datasets (which assume frontal shots) and degrades fine-tuned classifiers. The proposed mitigation:

1. Pre-filter for home presence and frame-fraction before classifying.
2. Build the few-shot exemplar set with yard-angle photos, not front-elevation photos — shifts the prompt's distribution to match deployment.
3. Demote high-confidence to medium-confidence for partial-visibility cases.
4. Allow the user to upload a separate front-elevation photo as a reference (this is the multi-image references feature — see Section C below).

Full detail in home-style-classification.md § A.5.

Section B: Existing Landscape Style Classification

The maturity gap. Where home-style classification has decades of prior art, landscape-style classification has almost none. The closest published work is the 2026 Nature Scientific Reports paper on garden design language, which built a 2,000-image dataset with semantic segmentation and aesthetic labels — research artifact, not a shipped classifier. The vegetation-classification literature (UAV multispectral, urban garden monitoring) solves a different problem (species / land-cover types from aerial imagery, not "is this a Cottage-leaning yard"). There is no off-the-shelf pretrained landscape-style classifier we can buy.

The proposed approach. Same architecture as home-style, different output. The vision LLM returns a soft membership distribution across the 12 design systems plus a "what's worth keeping" sub- classification. Two outputs from one vision call:

{
  "landscape_style_distribution": {
    "Cottage": 0.42, "Naturalistic": 0.21, "Rustic": 0.15,
    "Bare lawn": 0.10, "Modern Farmhouse": 0.08, "Other": 0.04
  },
  "primary_label": "Cottage-leaning",
  "keep_assessment": {
    "mature_trees": { "present": true, "remove_recommendation": "keep" },
    "hardscape": { "present": true, "remove_recommendation": "consider" },
    "lawn": { "present": true, "fraction_of_frame": 0.45 },
    "overgrown_or_dead": { "present": true }
  }
}

A soft distribution is more useful than a hard label. The 12-system label set is a continuous aesthetic space, not a discrete one. Real yards are mixed. Surfacing "Cottage-leaning, with Naturalistic elements" lets the picker show two relevant systems with ribbons rather than auto-locking one. The Modsy precedent is the closest analogue — its style quiz produces a "blended result that's incredibly personal" (Modsy quiz) rather than a single label.

Where this earns its keep is Step 1 (Prepare site). Without existing-landscape classification, the "Prepare the site" step is a generic "click things to remove" prompt. With the classification, it becomes a tailored checklist:

"Looks like a Cottage-leaning yard with mature trees and overgrown shrubs. Want to keep the trees and clear the shrubs? You can adjust below."

[✓ Keep mature trees] [✗ Remove overgrown shrubs] [✓ Keep stone path]

This is the single biggest UX win of the two classifications — it converts an open canvas into a one-click-corrected directed checklist. Detail in home-style-classification.md § B.4.

Validation. Because there is no public benchmark, validation is internal. Two parallel sets:

1. Style validation — ~100 hand-tagged real yard photos with primary style. Target: primary-label accuracy >70%, primary + secondary accuracy >85%.
2. Keep/remove validation — ~50 photos with hand-annotated keep / remove decisions per visible feature. Target: per-feature accuracy

   80%.

These thresholds are deliberately generous because the downstream UI forgives errors — wrong defaults are one-click corrections, not blocking failures.

Section C: Multi-Image and Reference-Image Patterns in Adjacent Tools

The brief introduces a References section with three subtypes (alternate base view, property reference, aspirational) plus the Property Design Profile entity. The deepest analogues are interior-design tools (Modsy, Havenly, Decorilla, Houzz, Pinterest) and frontier image-gen tools (Adobe Firefly, Midjourney, ReRoom, Spacely). Detailed eight-tool scan in multi-image-design-patterns.md.

Pattern 1 — Multi-image is normal, not exceptional. Every mature tool supports multiple references per project. The hard question is how to surface them in a panel-constrained UI, not whether to have them. Recommended UI for SimplyScapes' ~300px panel:

• Collapsible References section. Defaults to collapsed when no references attached.
• Three subtype-labeled slots (alternate views, property, aspirational). Each shows a horizontal scroll of small thumbnails (40-60px square) with a "+" tile to add. Match Adobe Firefly's typed-slots pattern (Style + Structure as separate first-class affordances).
• Soft cap of ~6 per subtype, ~10 total per generation. Cost and latency are the natural backpressure (each active reference adds ~25% to generation latency).
• Per-thumbnail on/off toggle for per-generation selection. Match the ReRoom Multi-Angle Rendering pattern — the user toggles which references are active for the next generation.

Pattern 2 — Profile is per-property / per-user, not per-area. Every tool with a profile (Modsy, Havenly, the new Spacely AI Design System) attaches it to the user or the property. The brief's Property Design Profile follows the dominant pattern. Spacely's Q1 2026 launch of an "AI Design System" feature — a property-level style collection that propagates across rooms — directly validates this entity for the adjacent interior-design vertical.

Pattern 3 — Soft profile output, not hard label. Modsy's quiz returns a "blended result" — a soft distribution rather than one style. Havenly's returns a single label. The Modsy approach is strictly better for prompt construction (richer signal) and matches the soft-distribution recommendation in Section B above. Use Modsy's pattern.

Pattern 4 — Override patterns are immature; closest prior art is outside design tools. Modsy/Havenly handled per-room override through human designer conversation, not UI. Spacely's AI Design System propagation does not yet expose a clear override UI. Houzz's ideabook system avoids the problem by having no profile at all. The closest UI pattern is Figma component overrides + CSS cascade UIs: an inheritance indicator at the top of the panel, an explicit override affordance on the system pill, and a "propagate to property" reverse path. See multi-image-design-patterns.md § G.3 for the full pattern.

Pattern 5 — Tagging references by intent is a Houzz-only pattern, v2-priority. Only Houzz lets users tag references by intent (color, layout, material) via the Ideabook structure. v1 ships untagged references; v2 adds optional tag chips per reference. Adoption likely <20% but power-user value is high.

Implications for the SimplyScapes data model:

The brief's References subsystem changes the prompt-construction layer. Each generation takes an N-tuple of input images (base + N references) plus the system fragment plus the step prompt. The spec must define:

1. The per-generation reference budget (cost / latency tradeoff). Recommendation: soft cap at 6 active references; warn the user above this; hard cap at 10.
2. How reference intent is described to the model (v1: implicit "consider this aesthetic"; v2: explicit tag-driven prompt construction).
3. How user selection of references is persisted on the resulting generation history entry. Recommendation: sticky across generations, with explicit "use different references for this generation" override.
4. The vision-LLM call extracts a structured summary from each reference (color palette + structural notes + style tags) at upload time, cached, and includes the summary — not raw bytes — in the downstream generation prompt to keep latency manageable.

Section D: Synthesis — Implications for SimplyScapes

The three deltas converge on a single technical pattern: Gemini 2.5 Pro Vision is the workhorse for everything outside the core image- generation call. Specifically:

• Site analysis call (one-time per property) returns: yard type, orientation, climate zone, size, removable objects, home architectural style + confidence + alternates, existing landscape style distribution + confidence, and the keep assessment structured object. This is one Gemini Vision call with a structured-output schema covering all fields.
• Reference image upload (per upload) triggers a Gemini Vision call that extracts color palette, structural notes, and style tags. The summary is cached on the reference entity and re-used at generation time. No re-extraction per generation.
• Per-generation reference summary is concatenated into the prompt-construction layer alongside the design-system fragment and the step prompt. Raw image bytes are not in the generation prompt — the structured summary is. This keeps generation latency bounded.

The combined classification reshapes Step 1 (Prepare site). Today's Step 1 is generic "click things to remove." With the classifications, it's a directed checklist: pre-populated keep / remove defaults from the keep_assessment, plus an existing-landscape-style ribbon ("Looks like a Cottage-leaning yard"). One-click corrections. This is the single biggest UX win of the entire delta round, and it arrives "for free" once the vision pipeline returns the structured output.

The Property Design Profile is the gravity well. Three of the four research deltas (auto-classification, profile inheritance, multi- image references) converge on the same entity:

• Auto-classification populates profile fields (home-style, existing- landscape distribution) at first site analysis.
• Profile inheritance means a property's design system, questionnaire answers, and home-style are read once and re-used across N visual designs (front yard, backyard, parkstrip).
• Property-level references (front-yard hardscape image, aspirational uploads) live on the profile and are selectable from any visual design on that property.

The profile is the data-layer fulcrum that makes the panel UX make sense. Without it, every visual design re-asks the same questions. With it, the user picks a system once and propagates everywhere — with per-area override available as a clean break-link operation borrowed from Figma component patterns.

Engineering load distribution. The delta mostly adds vision-LLM prompt engineering and a Property Design Profile entity. It does not require new system dependencies. The ranking by load:

| Work item | Engineering load | Risk | |-----------|-----------------|------| | Property Design Profile entity (Hasura schema, inheritance resolution) | Medium | Schema migration of existing visual designs | | Auto-classification fields in site-analysis vision prompt | Low | Calibration data needed pre-launch | | References subsystem (uploads, three subtypes, generation budget) | Medium | Cost/latency sensitivity of reference budget | | Inheritance / override UI inside ~300px | Medium-High | Density risk in the panel | | "What to keep" Step 1 directed checklist | Low | Falls out of the keep_assessment field |

The biggest risk is panel density — fitting the inheritance breadcrumb, override affordances, references manager, system pill, and five-step indicator into ~300px of horizontal panel. Section E lays that out as the central UX-phase open question.

Section E: Open Questions for UX Phase

The delta round answered four research questions (5–8 in the revised brief) at the research level. The UX phase has more granular questions to resolve in collaboration with Dan:

Auto-classification UX questions:

1. Confidence-threshold values. The 0.80 / 0.50 thresholds in Section A are starting estimates. Real numbers come from a labeled validation set. Should the UX flow accommodate either path ("confirm" must always be unobtrusive) or be hard-tuned to the final numbers? Probably the former.
2. Soft confirm UX inside ~300px. Chip menu (top-3 alternates), full dropdown (all 12), or modal sheet with picker grid? Trade-off between density and discoverability. Probably chip menu, with "see all" expanding to a sheet.
3. Existing-landscape distribution surfacing. Show the soft distribution to the user, or pick the primary and treat the rest as "alternate vibes"? Non-designers don't read distributions — probably the latter.
4. Recommendation persistence. When a user picks a non-recommended design system, does the "Recommended for your home" ribbon fade, stay, or flip into a different state ("Different from your home, but you can pull it off")?

Property Design Profile inheritance UX questions:

5. Inheritance breadcrumb design. A small "Backyard ← Property (Cottage)" breadcrumb at the panel top? Or a passive label on the system pill? Or both (breadcrumb when on a non-default area, label on the pill)?
6. Override affordance density. The system pill currently shows the active system. Adding a "change for area only / change for property" two-path menu is more density. Test in wireframe.
7. "Propagate to property" reverse path UX. When a user changes the backyard's system, when (and how) do we offer "make this the property default?" Probably one-tap follow-up after the change is committed, dismissable.

Multi-image references UX questions:

8. References-section default state. Collapsed when empty, expanded when references attached, or always collapsed with a count badge? Wireframe iteration.
9. Per-generation reference selection. Sticky across generations with explicit "use different references for this generation" override, or per-generation by default? Probably sticky.
10. Promoting a generation reference to a property reference. When the user attaches an inspiration image to a single generation, offer "save this to property references" as a one-tap follow-up? Probably yes.
11. Reference upload affordance. Drag-and-drop into the panel, click-to-upload button, or both? Mobile pattern?

Cross-cutting panel-density question:

12. Can the panel hold all of this? At ~300px wide, the panel needs the system pill, step indicator, drag-and-drop staging area, References section, generation history, and prompt input — plus inheritance breadcrumb and override affordances. The UX phase's headline deliverable is a wireframe set proving the density works. Probable patterns: accordion-collapsed sections, panel-internal modals/sheets for transient editors (questionnaire, Profile editor, References manager), and aggressive use of icon-with-text-label compactness.

These questions go to the UX Designer for the next round. home-style-classification.md § D and multi-image-design-patterns.md § I mirror these in their respective subdomains.

---

Recommended Next Steps

1. Run /ss-product spec to write the technical specification for Phase 1 implementation (two-phase routing + credit system + conversation persistence).
2. Run /ss-legal disclosure to generate a defensive publication covering: (a) conversational landscape design with domain object libraries, (b) intent routing in design-specific multimodal interfaces, (c) proactive clarification in generative design tools, (d) markup-guided generation for spatial design.
3. Flag for IP counsel review: Home Outside US12518067B2 (AI landscape design generation) and Adobe US11972757B2 (conversational image editing). Both are granted and active — architectural differentiation strategy should be validated.
4. A/B test Gemini 2.5 Flash Image vs. 3.1 Flash Image with 20-30 real landscape design prompts to validate quality/cost tradeoffs.
5. Prototype system prompts using the templates in Section 5 with Gemini API Playground to validate intent routing accuracy before building the full pipeline.
6. Migrate from Gemini 2.0 Flash to 2.5 Flash before the June 1, 2026 deprecation deadline.
7. Establish quarterly patent watch for Adobe, Google, Baidu, ByteDance, and Home Outside in CPC G06T11/00, G06F30/13.

Sources

| # | Type | Reference | URL | |---|------|-----------|-----| | 1 | API Docs | Gemini 2.5 Flash API — Function Calling | https://ai.google.dev/gemini-api/docs/function-calling | | 2 | API Docs | Gemini 2.5 Flash API — Image Generation | https://ai.google.dev/gemini-api/docs/image-generation | | 3 | API Docs | Gemini API Pricing | https://ai.google.dev/gemini-api/docs/pricing | | 4 | Paper | RouteLLM: Learning to Route LLMs with Preference Data (2024) | https://arxiv.org/abs/2406.18665 | | 5 | Paper | DialogGen: Multi-modal Dialogue for Multi-turn T2I (2024) | https://arxiv.org/abs/2403.08857 | | 6 | Paper | Talk2Image: Multi-Agent Multi-Turn Image Editing (2025) | https://arxiv.org/abs/2508.06916 | | 7 | Paper | TDRI: Two-Phase Dialogue Refinement for Interactive Generation (2025) | https://arxiv.org/abs/2503.17669 | | 8 | Paper | Proactive Agents for Multi-Turn T2I Under Uncertainty (2024) | https://arxiv.org/abs/2412.06771 | | 9 | Paper | SmartEdit: Complex Instruction-based Image Editing (CVPR 2024) | https://openaccess.thecvf.com/CVPR2024 | | 10 | Paper | BrushEdit: All-in-One Image Inpainting and Editing (2024) | https://arxiv.org/abs/2412.10316 | | 11 | Paper | ToolACE: Winning the Points of LLM Function Calling (ICLR 2025) | https://arxiv.org/abs/2409.00920 | | 12 | Benchmark | BFCL v3: Berkeley Function-Calling Leaderboard | https://openreview.net/forum?id=2GmDdhBdDk | | 13 | Paper | Training-Free Sketch-Guided Diffusion (2024) | https://arxiv.org/abs/2409.00313 | | 14 | Paper | AIdeation: Human-AI Collaborative Ideation (CHI 2025) | https://arxiv.org/abs/2502.14747 | | 15 | Paper | Effects of GenAI on Design Fixation (CHI 2024) | https://dl.acm.org/doi/10.1145/3613904.3642919 | | 16 | Paper | Multi-modal Intent Recognition Survey (EMNLP 2025) | https://aclanthology.org/2025.findings-emnlp.823 | | 17 | Paper | Improving LLM Function Calling via Structured Templates (EMNLP 2025) | https://arxiv.org/abs/2509.18076 | | 18 | Paper | CVPR 2024 Instruction-guided Editing Winning Solution | https://arxiv.org/abs/2407.13139 | | 19 | Industry | 2025 State of SaaS Pricing (Growth Unhinged) | https://www.growthunhinged.com/p/2025-state-of-saas-pricing-changes | | 20 | Industry | Rise of AI Credits (Metronome) | https://metronome.com/blog/the-rise-of-ai-credits | | 21 | Industry | AI Pricing 2025 Field Report (Metronome) | https://metronome.com/blog/ai-pricing-in-practice-2025-field-report | | 22 | Industry | Evolving AI SaaS Monetization (McKinsey) | https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/upgrading-software-business-models-to-thrive-in-the-ai-era | | 23 | SDK | Vercel AI SDK 6 — Tool Calling | https://ai-sdk.dev/docs/ai-sdk-core/tools-and-tool-calling | | 24 | Patent | US11972757B2 — Adobe Conversational Image Editing | https://patents.google.com/patent/US11972757B2 | | 25 | Patent | US12518067B2 — Home Outside AI Landscape Design | https://patents.google.com/patent/US12518067B2 | | 26 | Patent | US10579737B2 — Adobe NL Image Editing Framework | https://patents.google.com/patent/US10579737B2 | | 27 | Patent | US20230230198A1 — Adobe TiGAN Interactive Editing | https://patents.google.com/patent/US20230230198A1 | | 28 | Patent | US11983806B1 — OpenAI Inpainting/Outpainting | https://patents.google.com/patent/US11983806B1 | | 29 | Patent | US12039431B1 — OpenAI Visual Annotation Interaction | https://patents.google.com/patent/US12039431B1 | | 30 | Patent | US20250111139A1 — Adobe Design Document from Text | https://patents.google.com/patent/US20250111139A1 | | 31 | Patent | WO2024158398A1 — AI Drawing with Clarifying Questions | https://patents.google.com/patent/WO2024158398A1 | | 32 | Defensive | TDCommons — 22 publications on AI creative tools (2022-2026) | https://www.tdcommons.org/ | | 33 | Product | LeanScaper — Landscape Business Platform | https://www.leanscaper.com | | 34 | Product | iScape — Landscape Design App | https://www.iscapeit.com | | 35 | Product | PRO Landscape — Professional Design Software | https://www.prolandscape.com | | 36 | Product | Yardzen — Online Landscape Design | https://yardzen.com | | 37 | Product | Planter — Garden Planning App | https://planter.garden | | 38 | Open Source | RouteLLM (GitHub) | https://github.com/lm-sys/RouteLLM | | 39 | Open Source | BrushNet (GitHub) | https://github.com/TencentARC/BrushNet | | 40 | Open Source | Gorilla (GitHub) | https://github.com/ShishirPatil/gorilla | | 41 | Open Source | Open Pencil (GitHub) | https://github.com/open-pencil/open-pencil | | 42 | Open Source | LangGraph (GitHub) | https://github.com/langchain-ai/langgraph | | 43 | Report | WIPO GenAI Patent Landscape Report 2024 | https://www.wipo.int/web-publications/patent-landscape-report-generative-artificial-intelligence-genai | | 44 | Standards | ASLA Documentation Standards (Part V) | https://www.asla.org/2016awards/171889.html | | 45 | Extension | NC State Gardener Handbook — Landscape Design (Part V) | https://content.ces.ncsu.edu/extension-gardener-handbook/19-landscape-design | | 46 | Extension | UF/IFAS Basic Principles of Landscape Design (Part V) | https://ask.ifas.ufl.edu/publication/MG086 | | 47 | Industry | Loft Six Four — Landscape Architecture Process (Part V) | https://loftsixfour.com/blog/a-walk-through-the-landscape-architecture-process-what-developers-can-expect/ | | 48 | Industry | Verdance — Preliminary Design Phase (Part V) | https://www.verdancedesign.com/blog/landscape-architecture-process-preliminary-design | | 49 | Help Doc | Claude Design — Set up your design system (Part V) | https://support.claude.com/en/articles/14604397-set-up-your-design-system-in-claude-design | | 50 | Help Doc | Claude Design — Get started (Part V) | https://support.claude.com/en/articles/14604416-get-started-with-claude-design | | 51 | Press | Anthropic — Introducing Claude Design (Part V) | https://www.anthropic.com/news/claude-design-anthropic-labs | | 52 | Review | DataCamp — What Is Claude Design? (Part V) | https://www.datacamp.com/blog/claude-design | | 53 | Review | Build Fast with AI — Claude Design Guide 2026 (Part V) | https://www.buildfastwithai.com/blogs/claude-design-anthropic-guide-2026 | | 54 | Product | Spacely AI — Preset/Prompt Mode (Part V) | https://www.spacely.ai/ | | 55 | Process | Yardzen — How It Works (Part V) | https://yardzen.com/how-it-works | | 56 | Review | Yardzen Review (Lela Burris) (Part V) | https://www.lelaburris.com/yardzen-review/ | | 57 | Review | Modsy Review (Decorilla) (Part V) | https://www.decorilla.com/online-decorating/modsy-review/ | | 58 | Quiz | Havenly Style Quiz (Part V) | https://havenly.com/interior-design-style-quiz | | 59 | Product | Higharc Studio (Part V) | https://www.higharc.com/product/studio | | 60 | Product | Maket — AI Floor Plan Creation (Part V) | https://www.maket.ai/ | | 61 | Industry | Yardzen — Guide to Landscaping Styles (Part V) | https://yardzen.com/yzblog/landscaping-ideas-for-every-style | | 62 | Industry | Lamacchia — Popular Landscape Design Styles (Part V) | https://lamacchialandscapeco.com/landscape-design-styles/ | | 63 | Industry | Bullard Bollards — 7 Landscape Design Principles (Part V) | https://bullardbollards.com/7-landscape-design-principles-complete-guide-for-beautiful-outdoor-spaces/ | | 64 | Reference | The Landscape Library — Plant Symbols (Part V) | https://www.thelandscapelibrary.academy/blog/plant-symbols-for-landscape-design | | 65 | Paper | Xu et al., "Architectural Style Classification Using Multinomial Latent Logistic Regression" (ECCV 2014) (Part VI) | https://link.springer.com/chapter/10.1007/978-3-319-10590-1_39 | | 66 | Paper | "Architectural style classification based on CNN and channel-spatial attention" (Springer 2022) (Part VI) | https://link.springer.com/article/10.1007/s11760-022-02208-0 | | 67 | Project | McIntire — "Building an image recognition model to classify home types" (Part VI) | https://kurtmcintire.com/image-recognition-homes/ | | 68 | Project | Stanford CS231n — "Classifying U.S. Houses by Architectural Style Using CNNs" (Part VI) | https://cs231n.stanford.edu/reports/2017/pdfs/126.pdf | | 69 | Dataset | Architectural Styles dataset (Kaggle / Xu) (Part VI) | https://www.kaggle.com/datasets/dumitrux/architectural-styles-dataset | | 70 | Dataset | Roboflow ArchiStyles (Part VI) | https://universe.roboflow.com/mipt4/archistyles | | 71 | Paper | "Decoding garden design language via semantic segmentation" (Nature Sci Rep 2026) (Part VI) | https://www.nature.com/articles/s41598-026-46120-w | | 72 | Paper | "Identifying and Monitoring Gardens in Urban Areas Using Aerial and Satellite Imagery" (MDPI 2023) (Part VI) | https://www.mdpi.com/2072-4292/15/16/4053 | | 73 | Paper | "Utilizing active learning and attention-CNN to classify vegetation" (Nature Sci Rep 2024) (Part VI) | https://www.nature.com/articles/s41598-024-82248-3 | | 74 | Benchmark | LM Council — Vision benchmark comparison May 2026 (Part VI) | https://lmcouncil.ai/benchmarks | | 75 | Paper | Brain MRI sequence classification — GPT-4o vs Claude 4 vs Gemini 2.5 Pro (Part VI) | https://pmc.ncbi.nlm.nih.gov/articles/PMC12345967/ | | 76 | Paper | "Zero-Shot Fine-Grained Image Classification Using Large Vision-Language Models" (arXiv 2510.03903) (Part VI) | https://arxiv.org/abs/2510.03903 | | 77 | Paper | "GPT4Vis: What Can GPT-4 Do for Zero-shot Visual Recognition?" (Part VI) | https://arxiv.org/html/2311.15732v2 | | 78 | Article | Modexa — "The Confidence UI Pattern That Users Actually Trust" (Part VI) | https://medium.com/@Modexa/the-confidence-ui-pattern-that-users-actually-trust-ff27e1a8a956 | | 79 | Article | Orkes — "Human-in-the-Loop in Agentic Workflows" (Part VI) | https://orkes.io/blog/human-in-the-loop/ | | 80 | Reference | LlamaIndex — Confidence Threshold (Part VI) | https://www.llamaindex.ai/glossary/what-is-confidence-threshold | | 81 | Docs | Gemini API — Few-Shot Examples (Part VI) | https://cloud.google.com/vertex-ai/generative-ai/docs/learn/prompts/few-shot-examples | | 82 | Docs | Roboflow — Prompting Tips for LLMs with Vision (Part VI) | https://blog.roboflow.com/prompting-tips-for-large-language-models-with-vision/ | | 83 | Product | Modsy Style Quiz (Part VI) | https://www.modsy.com/design-style-quiz/ | | 84 | Industry | Modsy alternatives 2026 (Remodel AI) (Part VI) | https://www.remodelai.io/blog/modsy-alternatives | | 85 | Review | Havenly Review (Decorilla) (Part VI) | https://www.decorilla.com/online-decorating/havenly-review/ | | 86 | Review | Decorilla vs Havenly (Part VI) | https://www.decorilla.com/online-decorating/decorilla-vs-havenly-which-online-interior-design-service-is-right-for-you/ | | 87 | Review | Decorilla vs Modsy (Part VI) | https://www.decorilla.com/online-decorating/decorilla-vs-modsy/ | | 88 | Product | Houzz Quiz: Find Your Design Styles (Part VI) | https://www.houzz.com/ideabooks/houzz-quiz | | 89 | Tutorial | Houzz Ideabooks Tutorial (Jillian Lare) (Part VI) | https://jillianlare.com/houzz-ideabooks-tutorial/ | | 90 | Article | "Find Your Design Style" using Houzz Ideabooks (Part VI) | https://www.houzz.com/magazine/8-steps-to-finding-your-design-style-using-ideabooks-stsetivw-vs~178358057 | | 91 | Engineering | Pinterest Engineering — Building Pinterest Lens (Part VI) | https://medium.com/pinterest-engineering/building-pinterest-lens-a-real-world-visual-discovery-system-59812d8cbfbc | | 92 | Article | "Pinterest's Visual Lens: How computer vision explores your taste" (Part VI) | https://towardsdatascience.com/pinterests-visual-lens-how-computer-vision-explores-your-taste-5470f87502ad/ | | 93 | Docs | Pinterest Create — Boards (Part VI) | https://create.pinterest.com/product-features/how-to-create-boards/ | | 94 | Docs | Adobe Firefly — Reference images for styling (Part VI) | https://helpx.adobe.com/firefly/web/generate-images-with-text-to-image/customize-generated-images/reference-images-for-styling.html | | 95 | Docs | Adobe Firefly — Structure Reference (Part VI) | https://www.adobe.com/learn/firefly/web/generative-ai-structure-reference-image | | 96 | Docs | Midjourney — Style Reference (Part VI) | https://docs.midjourney.com/hc/en-us/articles/32180011136653-Style-Reference | | 97 | Docs | Midjourney — Omni-Reference (Part VI) | https://docs.midjourney.com/hc/en-us/articles/36285124473997-Omni-Reference | | 98 | Guide | Midjourney Omni-Reference Guide (ImaginePro) (Part VI) | https://www.imaginepro.ai/blog/2025/7/midjourney-omni-reference-guide | | 99 | Product | ReRoom AI (Part VI) | https://reroom.ai/ | | 100 | Product Update | Spacely AI Q1 2026 — AI Design System (Part VI) | https://resources.spacely.ai/whats-new-in-spacely-ai-your-complete-ai-design-system-has-arrived-q1-2026-edition/ |

---

Version History

| Date & Time (MT) | Author | Summary | |-------------------------|-------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | 2026-03-09 — MT | Research Analyst | Initial research report (Parts I–IV). Chat-API direction: intent routing, function calling, credit pricing, vertical and adjacent market scan, patent landscape, academic and open-source scan. | | 2026-05-01 04:45 PM MT | Research Analyst | Added Part V: UX Direction Pivot (Sections A–E). Three new supporting docs: landscape-designer-workflow.md, ai-design-tool-onboarding.md, landscape-design-systems.md. Sources table extended. Parts I–IV preserved as backend infrastructure reference. | | 2026-05-01 11:45 PM MT | Research Analyst | Added Part VI: Delta — Auto-Classification, Property Profile, Multi-Image (Sections A–E). Two new supporting docs: home-style-classification.md (covers home + landscape style classification) and multi-image-design-patterns.md (covers eight-tool reference-image scan). Recommends few-shot Gemini 2.5 Pro Vision over custom CNN for both classifications; soft-distribution output for landscape style; collapsible References section with three subtype slots; Figma-style override UI for Property Design Profile inheritance; "what to keep" Step 1 directed checklist as biggest UX win. Sources table extended (entries 65–100). Parts I–V preserved. |

Vertical Competitor Analysis: AI Features in Landscape Design

Date: 2026-03-09 Scope: AI capabilities, interaction models, and pricing across five landscape design vertical competitors Purpose: Inform SimplyScapes generative AI chat interface design decisions

---

Competitor Analysis

---

LeanScaper

How they approach it: LeanScaper is the first AI-powered platform purpose-built for the landscape industry, but it targets business operations rather than design visualization. The platform centers on "Lana," a conversational AI assistant trained specifically for landscaping businesses. Lana coordinates a suite of specialized agents (CFO Agent, CMO Agent, SOP Agent, Strategy Agent, Assessment Agent) that handle financial analysis, marketing campaigns, process documentation, and business planning. Users interact through natural language chat prompts like "Build me a marketing strategy for my ideal customer profile" or "Build my 2026 growth plan for my hardscape division." The mobile field app includes voice interaction for hands-free use during fieldwork, a Huddle Agent that records and processes crew meetings into structured notes and action items, and a Field Request Agent that automatically categorizes and routes requests to Kanban boards.

What works well:

• Chat-based conversational interface is natural and approachable for non-technical landscape contractors
• Domain-specific agents avoid the "blank prompt" problem by channeling users toward defined business functions
• Voice interaction from the field demonstrates practical understanding of how landscapers actually work
• Credit-based pricing aligns cost with value delivered rather than seat-based licensing
• Unlimited seats per plan removes adoption friction within organizations

Limitations / gaps:

• No landscape design visualization capabilities whatsoever -- this is purely a business operations platform
• No image generation, AR, or spatial design features
• The platform does not help homeowners or designers create or visualize landscape concepts
• Credit system may create anxiety about consumption, causing users to self-limit usage
• Relatively new to market (credit-based billing launched March 2026)

Technical approach: LLM-powered conversational AI with specialized agent routing. Knowledge base ("LeanDocs & Files") feeds context to agents. Voice-to-text for field use. No computer vision, image generation, or spatial computing components. Architecture appears to be a multi-agent orchestration layer on top of general-purpose LLMs, fine-tuned with landscape industry domain knowledge.

Pricing model:

• Free: $0/mo, 250 credits/mo, all features, unlimited users
• Core: $300/mo, 3,000 credits/mo (1-3 crew operations)
• Premium: $750/mo, 8,250 credits/mo (4-8 crews) -- most popular
• Max: $1,500/mo, 18,000 credits/mo (12+ crews)
• Top-up: $150 per 1,000 additional credits
• All plans include identical features; credits are the only differentiator

Key takeaway for SimplyScapes: LeanScaper validates the chat-based AI interaction model for the landscape vertical -- professionals are willing to engage with conversational AI for complex business tasks. Their agent-routing architecture (directing users to specialized agents by domain) is a pattern worth studying. However, the complete absence of design features means they occupy a non-overlapping space. SimplyScapes has an opportunity to bring the same conversational approachability to the design side of landscaping, which LeanScaper does not address at all.

---

iScape

How they approach it: iScape is the most-downloaded landscape design app (~4 million downloads, 4.6-star rating) and positions itself as the tool that lets homeowners and professionals design directly on photos of their own yard. The core technology is augmented reality (AR), not AI -- users place trees, shrubs, patios, and landscape elements using their smartphone camera in real-time 3D, or work in a 2D mode overlaying objects on photos. The interaction model is drag-and-drop: users browse a product catalog, select elements, and position them on their yard image. Professionals use the Pro tier to generate PDF proposals with pricing and business branding. iScape has announced an AI-powered design feature in beta that would help visualize outdoor designs using photos, but as of early 2026, AI is not a shipping core feature. App Store reviewers have specifically criticized the absence of AI capabilities that competing apps offer.

What works well:

• Photo-based design on your own yard is immediately understandable and builds trust in outcomes
• AR visualization removes ambiguity about scale, placement, and fit
• Massive product database (thousands of plants, hardscapes, products) gives specificity
• Professional proposal generation (PDF with pricing, branding) bridges design to sales
• Large existing user base provides distribution advantage for new feature adoption

Limitations / gaps:

• No shipping AI features -- AR is the visualization technology, not generative AI
• Users must manually browse catalogs and place every element; no automated design generation
• No chat interface or natural language interaction
• App Store reviews report crashes, slow loading, limited object manipulation (no tilt/rotate)
• Premium pricing ($29.99/mo or $299.99/yr) is high relative to the feature set for homeowners
• No intelligence around plant selection, climate zone recommendations, or companion planting

Technical approach: ARKit/ARCore-based spatial overlay on camera feed or static photos. Object library with pre-modeled 3D assets. Traditional UI (drag-and-drop, toolbars) rather than AI-driven workflows. The announced AI beta is uncharacterized technically but likely involves image generation given the "visualize designs using photos" description.

Pricing model:

• Free: Limited database access, basic 2D/3D design (iOS only)
• Plus: $14.99/mo (intermediate tier)
• Pro: $29.99/mo or $299.99/yr -- full database, proposals, custom uploads (most popular)
• Enterprise: Custom pricing, multi-user licenses, premium support, integrations
• No AI features are currently gated behind any tier

Key takeaway for SimplyScapes: iScape proves that photo-based design on your own yard is the interaction model homeowners want -- they do not want to design in abstract. However, their lack of AI creates a large opening. Users still must do all the creative work manually. SimplyScapes can leapfrog by combining the "design on your own photo" paradigm with generative AI that proposes complete designs, handles plant selection intelligently, and engages through conversation rather than catalog browsing. The gap between iScape's manual drag-and-drop model and a conversational AI design partner is the core opportunity.

---

PRO Landscape+ (Drafix Software)

How they approach it: PRO Landscape+ is the most AI-forward professional landscape design tool on the market as of early 2026. Drafix Software has aggressively integrated AI into a comprehensive desktop platform that combines photo imaging, CAD, 3D rendering, and proposal generation. Their AI suite includes five distinct tools: (1) AI Outdoor Living Designer -- upload a customer photo, define the project area, and generate complete outdoor living concepts including paver selections from 1,000+ manufacturer patterns, pergolas, outdoor kitchens, and fire features; (2) AI Paver Tool -- apply realistic paver layouts to photos with perspective adjustment and multi-pattern preview; (3) AI Eraser -- instantly remove existing landscaping from photos; (4) AI Cutout Tool -- precise object extraction for compositing; (5) Ask Wayne -- an AI help assistant (chatbot) for navigating the software. The interaction model is button-driven within a traditional desktop application: users click tools, define areas, and let AI generate or modify within bounded parameters. The key differentiator is that AI outputs feed directly into CAD for scaled drawings and material takeoffs, making AI-generated concepts immediately actionable for construction.

What works well:

• AI is deeply integrated into an established professional workflow, not bolted on
• AI-to-CAD pipeline (concept to scaled drawing to material takeoff to proposal) is uniquely complete
• Manufacturer-specific product libraries (1,000+ real paver patterns) ground AI outputs in purchasable reality
• The eraser and cutout tools address real pain points in photo-based design workflows
• Ask Wayne chatbot lowers the learning curve for a feature-rich desktop application

Limitations / gaps:

• Windows-only desktop application -- no mobile, no web, no cross-platform access
• Professional-only pricing and complexity; not accessible to homeowners
• AI interaction is tool-based (click buttons, define areas), not conversational or generative in the open-ended sense
• No natural language design requests (cannot say "design me a Mediterranean patio")
• Single-user license per computer; no collaborative or cloud-based workflows
• Ask Wayne appears limited to software help rather than design guidance

Technical approach: Desktop Windows application with integrated AI modules for image manipulation (eraser, cutout), generative concept creation (outdoor living designer), and pattern application (paver tool). Likely uses a combination of image segmentation models, inpainting/outpainting for the eraser, and template-based generation with perspective warping for paver layouts. The outdoor living designer appears to use constrained generation within defined project areas rather than unconstrained image generation. Ask Wayne is likely an LLM-based support chatbot scoped to software documentation.

Pricing model:

• Value: $900/yr ($75/mo equivalent, billed annually, saves $180)
• Flexibility: $90/mo (month-to-month)
• Both plans include all features, updates, support, training, and mobile companion app
• Single user, single computer license
• AI features are included in all subscriptions -- not separately gated

Key takeaway for SimplyScapes: PRO Landscape+ demonstrates the highest current bar for AI in landscape design and proves that professionals will adopt AI tools when they fit into existing workflows. Their AI-to-CAD-to-proposal pipeline is the gold standard for professional output. However, their desktop-only, Windows-only, professional-only positioning leaves the entire homeowner and mobile-first market unaddressed. SimplyScapes can bring comparable AI intelligence to a conversational, mobile-first, cloud-based experience that serves both homeowners and professionals. The Ask Wayne chatbot is notably limited to software help -- a genuine design-focused chat AI would be significantly more valuable.

---

Yardzen

How they approach it: Yardzen operates as a hybrid AI + human design service rather than a software product. Their model has two layers: (1) YardAI, a free web tool where homeowners upload a yard photo, choose from 16 design aesthetics (Cottage, Contemporary, Mediterranean, Rustic, Southwestern, Traditional, Modern Boho, etc.), and receive instant AI-generated landscape concepts; and (2) paid design packages ($995-$1,995) where professional landscape designers create custom CAD plans, photorealistic renderings, plant/materials lists, and lighting plans with human revision cycles. YardAI is explicitly positioned as an "inspiration tool" and "starting point," not a replacement for human design. It is trained on 50,000+ real landscape designs created by Yardzen's professional designers. After AI-only competitors emerged in 2024-2025, Yardzen deliberately leaned into human touchpoints -- adding dedicated project managers, one-on-one kickoff calls, and cost advisors. This counter-positioning resulted in a 28% increase in website conversion and 94% of customers choosing project-manager-guided packages over cheaper options.

What works well:

• Free AI tool (YardAI) serves as top-of-funnel acquisition -- users experience instant value before paying
• Training on 50,000 real professional designs produces more architecturally aware results than generic AI
• 16 named design styles give users a vocabulary for expressing preferences without design expertise
• Integrated product specifications (Belgard, PebbleTec, Crate & Barrel) ground AI suggestions in purchasable products
• Human-AI hybrid model builds trust: AI inspires, humans refine and finalize
• Deliberate counter-positioning against AI-only competitors proved that consumers value human expertise

Limitations / gaps:

• YardAI generates inspiration images, not actionable plans -- no measurements, no plant lists, no materials takeoff
• AI-generated images can include fictional elements and unrealistic trees
• No interactive refinement of AI concepts (cannot say "move the tree left" or "make it more rustic")
• Paid design packages are expensive ($995-$1,995) and slow (multi-week turnaround with revision cycles)
• 7% revenue share fee on contractor projects through their partner program
• No self-service design tools -- gap between free AI inspiration and expensive human service

Technical approach: Image-to-image generative AI (likely diffusion-based) trained on proprietary dataset of 50,000+ professional landscape designs. The model conditions on uploaded photos and selected style tokens. Spatial understanding (architecture, elevation, layout) is achieved through training data curation rather than explicit 3D reasoning. Product integration appears to be post-generation tagging/recommendation rather than generation-time conditioning. YardAI runs as a standalone web app (ai.yardzen.com) separate from the core service platform.

Pricing model:

• YardAI: Free ($0), unlimited use, web-based
• Essential package: $995 (or 6x $165.83) -- 2D CAD plans, plant/materials list, no revisions
• Classic package: $1,395 (or 6x $232.50) -- adds photorealistic renders, furniture selections, 1 revision, project manager
• Signature package: $1,995 (or 6x $332.50) -- adds lighting plan, nighttime renders, 2 revisions, cost advisor
• Installment plans available (6-month split)

Key takeaway for SimplyScapes: Yardzen's two-layer model (free AI inspiration to paid human design) is the most strategically instructive competitor. Their YardAI proves that free AI-generated landscape concepts are a powerful acquisition tool, but their deliberate pivot toward human touchpoints reveals that consumers do not trust AI alone for high-stakes design decisions. SimplyScapes should study the "gap" in Yardzen's model: there is no middle tier between free AI inspiration images and $995+ human design packages. A conversational AI chat interface that offers iterative refinement, actionable plant lists, and budget-aware recommendations could occupy this profitable middle ground.

---

Planter

How they approach it: Planter is a focused garden planning app for vegetable gardens and raised beds, operating on a square-foot gardening model. The interaction model is purely visual drag-and-drop: users create a grid representing their garden bed, then drag plant icons onto squares. The app provides companion planting guidance (which plants work well together and which conflict), automatic spacing calculations (e.g., tomatoes occupy 4 squares, shallots occupy 1), planting calendars based on hardiness zone, and a database of 80+ base plants with 1,000+ varieties. There is no generative AI -- the "intelligence" is rule-based: when you place a plant, the app flags compatible and incompatible neighbors and suggests optimal timing. The only AI feature is a minor one: AI-generated icons for custom plant varieties (added September 2025). The app is praised for simplicity and focus, earning strong reviews from vegetable gardeners.

What works well:

• Extreme simplicity -- plan a garden in minutes with no learning curve
• Companion planting rules are immediately useful and educational
• Zone-based planting calendar personalizes timing recommendations
• Affordable pricing ($24.99/yr or $99.99 lifetime) makes it accessible
• Narrow focus on vegetable gardening means it does one thing very well
• Cross-platform (iOS, Android, web) with sync

Limitations / gaps:

• No AI beyond icon generation -- all planning intelligence is rule-based
• No landscape design capability (only vegetable/herb gardens in raised beds)
• No photo-based visualization of how the garden will look
• No chat or conversational interface
• No generative design suggestions ("design me a salsa garden" is not possible)
• Limited to square-foot gardening paradigm; does not handle ornamental or full-yard design
• No climate-aware plant recommendations beyond basic hardiness zone

Technical approach: Grid-based spatial planner with a plant database storing companion relationships, spacing requirements, and zone-specific planting windows. Rule engine evaluates plant adjacencies and flags conflicts. No machine learning, computer vision, or generative models. The AI icon generation (for custom plants) likely uses an image generation API but is peripheral to core functionality.

Pricing model:

• Free: 1 garden, calendar, custom plants
• Premium: $24.99/yr -- unlimited gardens, no ads, notes, custom backgrounds, web access
• Lifetime: $99.99 one-time -- all Premium features permanently
• No AI-specific pricing tier

Key takeaway for SimplyScapes: Planter demonstrates that even without AI, a well-designed companion planting engine with clear visual feedback creates genuine user delight. Their rule-based intelligence (companion/combative plant relationships, zone-based timing, spacing logic) represents baseline domain knowledge that any AI-powered landscape tool should embed. SimplyScapes should incorporate this kind of horticultural intelligence into its AI chat interface -- when a user asks for plant recommendations, the system should understand companion planting, spacing, zone compatibility, and seasonal timing as foundational knowledge, then layer generative design capabilities on top.

---

Vertical Market Patterns

1. AI adoption is bimodal. Competitors are either deeply invested in AI (PRO Landscape+, Yardzen, LeanScaper) or have essentially no AI features (iScape, Planter). There is no gradual middle ground -- companies either committed to AI as a core strategy or have not yet started.

2. Photo-based design is the baseline expectation. Every design-focused competitor (iScape, PRO Landscape+, Yardzen) uses the customer's own yard photo as the starting canvas. Abstract design tools with no connection to the user's actual space are not competitive.

3. AI is used for generation, not conversation. None of the competitors offer a true conversational AI interface for design. PRO Landscape+ uses button-driven AI tools. Yardzen uses style-selector-to-generation. LeanScaper has chat, but for business ops, not design. The conversational design partner is an unoccupied niche.

4. Professional and consumer markets remain separate. PRO Landscape+ serves professionals exclusively (Windows desktop, $900/yr). iScape bridges both but leans consumer. Yardzen is consumer-only with human designers. No single product serves both audiences with AI-powered design.

5. The funnel gap is consistent. Yardzen's free AI inspiration has no self-service upgrade path below $995. iScape's free tier has no AI at all. There is a consistent gap between "free/cheap exploration" and "professional output" that no competitor fills with AI.

6. Manufacturer product integration is emerging. Both PRO Landscape+ (1,000+ paver patterns) and Yardzen (Belgard, PebbleTec, Crate & Barrel) are integrating real purchasable products into AI outputs. This grounds designs in commercial reality and opens monetization through product partnerships.

7. Human-AI hybrid models outperform AI-only. Yardzen's deliberate pivot toward human touchpoints (project managers, kickoff calls) after AI-only competitors emerged -- resulting in 28% higher conversion -- signals that consumers want AI assistance but human validation for high-stakes outdoor renovation decisions.

8. Credit-based and usage-based pricing is emerging. LeanScaper's credit model (where different actions cost different amounts) represents a new pricing paradigm in the vertical, moving away from flat subscriptions toward value-aligned consumption pricing.

---

Vertical Market Gaps

1. No conversational AI design partner exists. No competitor offers a chat-based interface where homeowners can describe what they want in natural language and iteratively refine a landscape design through conversation. This is the single largest gap in the market.

2. No AI-powered middle tier between inspiration and professional design. Yardzen's model exposes a $995 gap between free AI inspiration images and human-designed plans. A product that generates actionable (not just inspirational) landscape plans through AI at $50-200 would address massive unmet demand.

3. No mobile-first AI design tool. PRO Landscape+ is Windows-only desktop. YardAI is a basic web tool. iScape is mobile but has no AI. No competitor combines mobile-first design with generative AI capabilities.

4. No AI-driven plant intelligence. Despite landscape design being fundamentally about plants, no competitor uses AI for intelligent plant selection that considers climate zone, soil type, sun exposure, water requirements, companion planting, seasonal interest, maintenance level, and budget simultaneously. Planter has rule-based companion planting; everyone else treats plants as catalog items.

5. No iterative AI refinement. YardAI generates a concept but cannot refine it ("make it more drought-tolerant," "swap the maple for something smaller"). PRO Landscape+ AI generates within constraints but does not accept natural language feedback. The ability to conversationally iterate on AI-generated designs is absent from every competitor.

6. No budget-aware AI design. No competitor's AI considers budget constraints when generating designs. Cost information exists (Yardzen has a cost advisor, PRO Landscape+ does material takeoffs) but is not integrated into the generative process. An AI that designs within a stated budget would be differentiated.

7. No cross-platform cloud collaboration. PRO Landscape+ is single-user, single-machine. iScape is single-device. Yardzen collaboration happens through the service team. No competitor offers real-time collaborative design editing with AI assistance across devices.

8. No AI that learns from the user's property. Despite every competitor starting from a user's photo, none build a persistent model of the property (existing plants, soil conditions, sun patterns, irrigation) that improves recommendations over time. Each design session starts from scratch.

9. No integration between business operations AI and design AI. LeanScaper does business ops. PRO Landscape+ does design. No product bridges both -- an AI that helps design the landscape AND generates the proposal, schedules the crew, and estimates the job would be uniquely comprehensive.

10. No AI-generated maintenance plans. Every competitor focuses on the design moment. None use AI to generate ongoing maintenance schedules, seasonal care instructions, or long-term landscape evolution plans based on the design they helped create.

Adjacent Market Analysis: AI Interaction Patterns, Credit Systems & Design Integration

Date: 2026-03-09 Type: Supporting Research — Competitive Intelligence Parent: generative-ai-chat-interface

---

Purpose

This analysis examines seven adjacent-market products that have built AI-powered creative tools with interaction patterns, credit systems, and design integration approaches relevant to SimplyScapes. Each product offers transferable lessons for building a generative AI chat interface for landscape design — even though none of them operate in the landscaping vertical.

---

Product Analyses

---

Canva AI (Magic Studio) — Democratized AI design with credit-gated features

Their solution: Canva Magic Studio bundles over 25 AI-powered tools directly into the design canvas, including Magic Design (prompt-to-layout), Magic Media (text-to-image using DALL-E and Imagen backends), Magic Write (AI copywriting), Magic Animate (one-click animation), and Magic Eraser. Rather than presenting AI as a separate mode, Canva distributes AI tools throughout the editing interface — users encounter them contextually right where they are working. The platform supports multimodal input (text prompts, image uploads, voice dictation) and outputs across text, image, and video. Users can select from modalities ("Design," "Image," "Doc," "Code," "Video clip") and fine-tune results by providing feedback, uploading reference images, or trying suggested prompts.

The pattern worth studying:

1. Contextual AI placement: AI tools are embedded at the point of need rather than siloed in a separate "AI mode." Users do not leave their workflow to access AI features.
2. Progressive disclosure of complexity: Free users get 50 AI image generations total; Pro users get 500/month. The system introduces AI capabilities gradually — users see what is possible, hit a soft limit, and are nudged toward paid tiers.
3. Multi-backend abstraction: Canva routes to OpenAI DALL-E, Google Imagen, or its own models behind a single "Magic Media" interface. Users never choose a model — they describe what they want and the platform picks the best backend.
4. Real-time usage tracking: As of March 2026, Canva introduced a real-time tracker so users can monitor their AI usage allowance within the app settings.
5. Suggested prompts and inspiration: When users open the image generation tool, they see suggested prompt ideas to reduce the blank-canvas problem.

How it could adapt to landscaping:

• Embed AI generation directly into the landscape design canvas — e.g., a user selects an empty yard area and an "AI Fill" option appears contextually.
• Use the multi-backend abstraction pattern to route landscaping prompts to the best model (e.g., a fine-tuned garden model for plant placement, a general model for hardscape rendering).
• Offer suggested prompts tuned to landscaping scenarios: "Modern xeriscaped front yard," "Shade garden under mature oaks," "Low-maintenance backyard with play area."
• Implement real-time credit tracking so landscaping professionals can monitor usage across their team.

What doesn't transfer:

• Canva's design paradigm is 2D/flat templates. Landscaping requires 3D spatial reasoning, plant growth simulation, and terrain awareness that template-based generation cannot handle.
• The "everything is a template" mental model breaks down when users need site-specific designs tied to actual property dimensions.
• Canva's AI is generic — it has no domain knowledge about hardiness zones, drainage, sun exposure, or plant compatibility.

Credit/pricing model:

• Free: 50 total AI generations (lifetime cap)
• Pro ($15/month): 500 AI generations per month
• Teams ($10/user/month, 3-user minimum): 500 AI generations per user per month
• Business (~$20/user/month): Same AI allowance plus Leonardo.ai integration and IP indemnification
• Credits are per-user, not pooled (except enterprise custom deals)
• No per-generation pricing — AI is bundled into the subscription tier
• Overage: users hit a wall and must upgrade; no pay-as-you-go option for extra credits

---

Figma AI — In-context AI assistance tightly coupled to the design canvas

Their solution: Figma has layered AI capabilities directly into the design tool rather than offering a separate AI product. Key features include: Figma Make (prompt-to-prototype generation that creates responsive layouts, components, and interactive prototypes from natural language), First Draft (AI-generated starting points for common UI patterns), AI image editing tools (erase object, isolate object, expand image — all operating directly on canvas elements without text prompts), and Code-to-Canvas (announced February 2026 with Anthropic, enabling developers to push Claude-generated UI code directly into Figma as editable design layers). Critically, Figma Make is not just a generator — users can start with a prompt, refine visually, adjust generated code, or re-prompt to explore new directions, all within a single workspace.

The pattern worth studying:

1. Bidirectional AI-canvas integration: AI does not just generate static outputs — it produces fully editable, structured design objects that integrate into the user's existing design system (colors, typography, components from their libraries).
2. Granular credit pricing by complexity: Simple AI actions (image edits) cost fewer credits than complex ones (full prototype generation at 100+ credits). This aligns cost with value delivered.
3. Non-prompt AI tools: The erase/isolate/expand image tools require no text prompts — they work through direct manipulation (select an area, click "erase"). This is important for users who find prompting difficult.
4. Design system awareness: Figma Make can import and respect existing design libraries, so AI output is brand-consistent rather than generic.
5. Iterative refinement loop: Users prompt, see a result, refine visually, re-prompt if needed — the AI output is a starting point, not a final product.

How it could adapt to landscaping:

• Build a landscape design canvas where AI-generated plans are fully editable objects (draggable plants, resizable hardscape elements, adjustable zones) rather than flat images.
• Offer direct-manipulation AI tools for landscape editing: select a garden bed area and click "fill with shade-tolerant perennials" without needing to write a prompt.
• Let users import their design preferences (favorite plant palettes, preferred styles, material choices) so AI output respects their established design language.
• Implement the iterative refinement loop: AI generates a base landscape plan, the user adjusts plant placement, AI fills in gaps or suggests complementary plantings.

What doesn't transfer:

• Figma's output is screen-based UI — it has no concept of physical space, elevation changes, or real-world constraints.
• The Code-to-Canvas pattern (converting code to visual design) does not have a direct landscaping equivalent, though the concept of "specification to visual" could apply.
• Figma's component library model assumes reusable, identical UI elements; landscaping elements are more variable (each tree grows differently, each site has unique conditions).

Credit/pricing model:

• Starter: 150 credits/day, 500/month maximum
• Professional ($5/editor/month): 3,000 credits/month per full seat
• Organization ($5/editor/month): 3,500 credits/month per full seat
• Enterprise ($5/editor/month): 4,250 credits/month per full seat
• Dev/Viewer seats: 500 credits/month on all paid plans
• Credit costs by action: Simple image edits ~few credits; Figma Make generation 30-100+ credits depending on complexity
• New (March 2026): AI credit subscriptions (shared team pool at better rates) and pay-as-you-go billing (Q2 2026)
• Credits reset monthly; enforcement of limits begins March 18, 2026

---

Midjourney — Conversational image generation with voice-driven iteration

Their solution: Midjourney is an AI image generation platform that originated on Discord and has since built a full web interface at midjourney.com. Its core innovation is the conversational, iterative generation workflow. In Conversational Mode, users describe ideas in natural language to an AI that writes optimized prompts on their behalf. Users can activate voice input (click microphone, speak, stop) for a hands-free experience. The platform recently introduced Draft Mode for rapid idea exploration and a unified web editor combining inpainting (brush-based region replacement), outpainting (canvas extension), smart select, layers, and a "Suggest Prompt" feature that reverse-engineers prompts from existing images. V7 added personalization — users rate ~200 images and the model adapts toward their aesthetic preferences.

The pattern worth studying:

1. Conversational-to-prompt translation: Users speak naturally ("a peaceful Japanese garden with a koi pond") and the AI translates this into an optimized generation prompt. This removes the "prompt engineering" barrier entirely.
2. Grid-based choice presentation: Midjourney generates 4 variations simultaneously (a 2x2 grid), letting users compare options at a glance before committing to upscale (U1-U4) or create variations (V1-V4).
3. Voice-driven creative workflow: The Draft Mode + voice input combination allows rapid ideation — talk, see results, tweak, re-roll — without typing.
4. Personalization through preference learning: By rating ~200 images, users train a personal style model. This is a passive way to capture design preferences without requiring users to articulate them.
5. Unified editor for post-generation refinement: The web editor brings inpainting, outpainting, layers, and re-prompting into a single view, so users can refine AI output without leaving the platform.
6. Suggest Prompt (reverse engineering): Users can upload an image and get an AI-generated prompt that would recreate it — useful for understanding and iterating on reference images.

How it could adapt to landscaping:

• Implement conversational prompt translation for landscape design: "I want a low-maintenance front yard with some color" becomes an optimized prompt with specific plant suggestions for the user's zone.
• Show 4 landscape plan variations side-by-side for comparison — different styles, plant palettes, or layout approaches for the same yard.
• Offer voice-driven design sessions where landscapers describe their vision while looking at the property, and the AI generates concepts in real time.
• Build preference learning into the onboarding: show users 50-100 landscape photos, have them rate preferences, and use this to bias all future AI output toward their aesthetic.
• Allow "Suggest Prompt" from existing landscape photos: upload a photo of a yard the user admires, and the AI describes the design elements to use as a starting point.

What doesn't transfer:

• Midjourney generates artistic images, not actionable plans. A beautiful rendering of a garden is not the same as a plantable design with species names, spacing, and installation instructions.
• The Discord-first heritage created a community-oriented workflow (public galleries, shared channels) that does not fit a B2B landscaping tool where designs are client-confidential.
• GPU time billing works for art generation but is hard to map to landscaping design tasks that have variable complexity (a 500 sq ft patio vs. a 5-acre estate).

Credit/pricing model:

• Basic ($10/month, $96/year): ~3.3 GPU hours/month of Fast generation
• Standard ($30/month, $288/year): ~15 GPU hours/month Fast + unlimited Relax Mode
• Pro ($60/month, $576/year): ~30 GPU hours/month Fast + unlimited Relax + Stealth Mode (private gallery)
• Mega ($120/month, $1,152/year): ~60 GPU hours/month Fast + unlimited Relax + Stealth Mode
• Billing is by GPU time, not per image — a standard 4-image grid uses ~1 minute of Fast time
• Relax Mode (Standard+ plans) offers unlimited generations at lower priority, with no Fast time consumption
• No free tier or trial as of 2025
• Annual billing gives ~20% discount

---

Adobe Firefly — Enterprise-grade AI generation with IP-safe credit pooling

Their solution: Adobe Firefly is Adobe's generative AI engine, integrated across Creative Cloud applications (Photoshop, Illustrator, Premiere Pro, After Effects) and available as a standalone web app. Its key differentiator is IP indemnification — Firefly models are trained exclusively on Adobe Stock, openly licensed content, and public domain material, making outputs commercially safe. The credit system is tiered by action complexity: standard image generation costs 1 credit, while premium video generation costs 20-100 credits per second. Enterprise customers can access Firefly Foundry for brand-specific model training on their own content, guidelines, and IP. Recent additions include Firefly Boards (collaborative AI generation workspace) and deeper video integration with Premiere Pro.

The pattern worth studying:

1. Tiered credit cost by output type: Standard image generation (1 credit) vs. premium video (20-100 credits/second) creates a natural value hierarchy where more expensive outputs cost more.
2. IP indemnification as a feature: Adobe guarantees commercial safety of AI outputs and provides legal protection. This is a premium feature that justifies higher pricing.
3. Deep integration into existing professional workflows: Firefly is not a standalone tool — it lives inside Photoshop's Generative Fill, Illustrator's vector generation, and Premiere's video editing. Professionals never leave their primary tool.
4. Enterprise credit pooling: While individual plans do not pool credits, enterprise customers can purchase shared credit pools, allowing teams to allocate AI resources based on project needs.
5. Brand-specific model training (Firefly Foundry): Enterprise customers can train custom models on their own content, ensuring AI output matches brand guidelines.
6. Unlimited standard generation promotions: Adobe periodically offers unlimited standard image generation to subscribers, using it as a growth lever and upsell mechanism.

How it could adapt to landscaping:

• Implement tiered credit costs: a quick "style preview" rendering might cost 1 credit, while a full 3D walkthrough video might cost 20+ credits.
• Offer IP indemnification for AI-generated landscape designs — guarantee that generated designs do not infringe on existing landscape architecture IP.
• Integrate AI generation directly into whatever design tool landscapers already use, rather than requiring them to switch to a separate AI tool.
• Enable landscape companies to train custom models on their portfolio of completed projects, so AI output matches their established design style.
• Offer enterprise credit pools so a landscape company can allocate AI credits across their design team based on project needs.

What doesn't transfer:

• Adobe's ecosystem lock-in strategy (Firefly works best inside Creative Cloud) may not apply — SimplyScapes is building a new platform, not extending an existing one.
• The IP indemnification model requires training exclusively on licensed content, which is expensive and may limit output diversity for landscaping use cases.
• Adobe's pricing assumes professional creative workers who already pay $55-80/month for Creative Cloud. Landscaping professionals have different software budgets.

Credit/pricing model:

• Firefly Standard ($9.99/month): 2,000 generative credits/month
• Firefly Pro ($29.99/month): 7,000 generative credits/month
• Firefly Premium ($199.99/month): 50,000 generative credits/month
• Creative Cloud All Apps ($59.99/month): Includes 4,000 credits/month (recently increased from 3,000)
• Credit costs by action: 1 credit per standard image generation; 20 credits/second for 1080p video; 100 credits/second for high-quality video; 5 credits/second for translation
• Enterprise (ETLA): 4,000-8,000 credits/user/month; optional credit pool add-on
• On-demand purchase: Available for extra credits beyond monthly allocation
• Credits are per-user, not pooled on individual and team plans; enterprise can optionally pool

---

Runway ML — Creative AI with transparent per-second API pricing

Their solution: Runway ML is a browser-based creative AI platform focused on video and image generation, offering models like Gen-4 (text/image-to-video), Gen-3 Alpha, and integrations with third-party models like Google Veo 3. The platform provides both a consumer-facing web interface and a developer API for embedding AI generation into custom applications. The API uses a simple, transparent pricing model: credits are purchased at $0.01/credit, and each model has a per-second credit cost (e.g., Gen-4 Video at 12 credits/second = $0.12/second). The web interface provides tools for text-to-video, image-to-video, video editing, and image manipulation, with Gen-4 achieving strong character consistency across scenes using reference images.

The pattern worth studying:

1. Transparent per-unit API pricing: $0.01/credit with published per-second costs per model. Developers can calculate exact costs before building. No hidden fees or complex tiering.
2. Tiered model quality/speed/cost: Gen-4 Turbo (5 credits/sec, fast, cheaper) vs. Gen-4 standard (12 credits/sec, higher quality) vs. Gen-4 Aleph (15 credits/sec, highest quality). Users choose their quality-cost tradeoff.
3. "Unlimited" with quality tradeoff: The Unlimited plan ($76/month) offers unlimited generations at "relaxed rate" (lower priority queue). This lets heavy users generate without anxiety while managing infrastructure costs.
4. Dual interface strategy: A polished web UI for creative professionals and a developer API for integration into custom workflows. Same underlying models, different access patterns.
5. Reference image consistency: Gen-4 uses reference images to maintain character/scene consistency across multiple generations — critical for any design workflow that requires coherent output across iterations.

How it could adapt to landscaping:

• Offer transparent per-generation pricing for the API, so landscape software companies can embed SimplyScapes AI generation at predictable costs.
• Provide tiered quality levels: "Quick Preview" (fast, low credit cost, low resolution) vs. "Client Presentation" (slower, higher cost, photorealistic rendering).
• Offer a "Relax Mode" equivalent for unlimited low-priority generation, useful for exploration and brainstorming phases of landscape design.
• Build both a web UI (for landscape designers using SimplyScapes directly) and an API (for integration into existing landscape design software).
• Use reference image consistency to maintain design coherence across multiple views of the same landscape (front view, side view, aerial view).

What doesn't transfer:

• Runway is optimized for video generation, which has a very different cost structure from landscape plan generation.
• The per-second billing model assumes time-based output (video). Landscape design output is area-based or element-based, requiring a different unit of measurement.
• Runway's creative user base (filmmakers, video editors) has different expectations and workflows than landscape professionals.

Credit/pricing model:

• Free: 125 credits (one-time, not recurring)
• Standard ($12/month): 625 credits/month
• Pro ($28/month): 2,250 credits/month
• Unlimited ($76/month): 2,250 credits/month + unlimited Relax Mode (lower priority)
• Enterprise: Custom pricing
• API: $0.01 per credit; Gen-4 Video = 12 credits/sec ($0.12/sec); Gen-4 Turbo = 5 credits/sec ($0.05/sec); Veo 3 = 40 credits/sec ($0.40/sec)
• Credits do not roll over; monthly reset
• No per-user model — credits are per-account/organization

---

ChatGPT (Image Generation) — Multi-turn conversational image editing

Their solution: OpenAI replaced DALL-E 3 with GPT-4o's native image generation capabilities in ChatGPT (March 2025-2026), representing an architectural shift from a separate image generation model to a unified model that can both converse and generate/edit images natively. The key breakthrough is multi-turn image editing: unlike DALL-E 3's regeneration-only approach, GPT-4o enables iterative refinement through conversation. Users can request specific changes ("move the tree to the left," "make the flowers more vibrant," "change the pathway material to flagstone") and the model modifies the existing image without starting from scratch. GPT-4o also resolved longstanding issues with text rendering in images and hand anatomy. Free users get 2-3 images per day; Plus subscribers ($20/month) get 50 images per 3-hour window.

The pattern worth studying:

1. Conversational image refinement: Users edit images through natural language dialogue rather than masking tools or parameter sliders. "Make the garden path wider" is more intuitive than drawing a selection and adjusting width.
2. Context-aware editing: The model understands the full conversation history, so it can make coherent changes across multiple editing rounds without losing the overall design intent.
3. Unified text + image model: There is no modal switch between "chat mode" and "image mode" — the same conversation can include questions, explanations, and image generation/editing seamlessly.
4. Low barrier to entry: Free tier access (2-3 images/day) lets users experience the capability before paying. The limit is tight enough to drive upgrades but generous enough for evaluation.
5. Iterative refinement without re-prompting from scratch: Users build on previous generations rather than starting over, which preserves design intent and reduces wasted credits.
6. Rolling window rate limits: The 50-images-per-3-hours model prevents abuse while allowing burst usage for active design sessions.

How it could adapt to landscaping:

• Build a conversational landscape design interface where users describe changes in natural language: "add a water feature near the patio," "replace the lawn with native grasses," "show me this design in fall colors."
• Maintain conversation context across a design session so the AI remembers the full design history and makes coherent incremental changes.
• Eliminate the modal split between "asking questions about plants" and "generating a design" — the same conversation should handle both seamlessly.
• Offer a free tier with limited daily generations to let homeowners experience AI landscape design before committing to a paid plan.
• Implement rolling window limits (rather than hard monthly caps) to support the bursty nature of design sessions.

What doesn't transfer:

• ChatGPT generates images as flat raster outputs with no structured data (no plant species names, no dimensions, no bill of materials). Landscape design needs structured output alongside the visual.
• The model has no domain-specific knowledge of plant compatibility, hardiness zones, sun requirements, soil types, or local regulations.
• Multi-turn editing works well for single images but does not scale to multi-view landscape plans (plan view, elevation, 3D walkthrough) that need to stay synchronized.
• ChatGPT's rate limiting model (per-user, per-time-window) does not account for the collaborative nature of landscape design where a designer and client may iterate together.

Credit/pricing model:

• Free: 2-3 images per day (24-hour rolling window)
• Plus ($20/month): 50 images per 3-hour rolling window
• Pro ($200/month): Higher limits, priority access
• API (GPT-4o image generation): Per-token pricing; image output tokens are priced at the standard GPT-4o output rate
• No explicit "credit" system — rate limits are time-based rather than credit-based
• Free tier is a true free tier (no trial expiration), providing ongoing low-volume access
• Image generation is bundled with chat — no separate image pricing

---

Vercel AI SDK — Provider-agnostic infrastructure for building AI interfaces

Their solution: The Vercel AI SDK is a free, open-source TypeScript toolkit for building AI-powered applications, primarily targeting Next.js. It is not an AI product itself but rather the infrastructure layer for building AI products. Key abstractions include: Provider abstraction (a unified interface across OpenAI, Anthropic, Google, Mistral, and self-hosted models — switch providers by changing one line of code), Streaming (Server-Sent Events streaming with React hooks that reduce boilerplate from 200-300 lines to 10-20 lines), Tool/Function calling (type-safe tool definitions with Zod schemas, automatic conversation management, and multi-step agent loops), Structured output (type-safe schema enforcement for AI responses), and the new Agent abstraction (AI SDK 6, announced 2025, introducing a reusable Agent interface with tools, instructions, and type-safe UI streaming). The SDK is free and works with any hosting provider — it is not locked to Vercel.

The pattern worth studying:

1. Provider abstraction eliminates vendor lock-in: A unified interface means the application can switch between OpenAI, Anthropic, Google, or a custom model without rewriting UI code. This is critical for a startup that may need to change providers as models improve or costs change.
2. Streaming-first architecture: streamText for user-facing features (tokens appear as generated) vs. generateText for background tasks. This distinction is important for landscape design where some operations are interactive (chat) and others are background (rendering).
3. Type-safe tool calling: Tools are defined with Zod schemas, ensuring the AI can only call functions with valid parameters. For landscape design, this means AI-suggested plant placements would be validated against the schema before being applied to the canvas.
4. Agent loop with automatic conversation management: The SDK handles the full tool execution loop — LLM decides to call a tool, SDK executes it, result is appended to conversation, new generation triggered — until a text response is produced. This pattern enables complex multi-step design operations.
5. Speech integration (AI SDK 5+): Unified speech generation and transcription interface, enabling voice-driven design workflows.
6. Framework-agnostic but Next.js-optimized: Works with any React framework but provides optimized patterns for Next.js App Router, which is the SimplyScapes stack.

How it could adapt to landscaping:

• Use the provider abstraction to route different landscape AI tasks to different models: plant identification to a vision model, design generation to an image model, conversation to a language model — all through a single SDK interface.
• Implement streaming for interactive design features (show the landscape plan rendering progressively) and background generation for complex operations (full 3D rendering).
• Define landscape-specific tools with Zod schemas: addPlant({ species: z.string(), position: z.object({x, y}), zone: z.number() }) — the AI can only suggest valid plant placements.
• Build an agent loop for multi-step landscape design: user says "design a low-maintenance backyard," agent calls plant database tool, checks zone compatibility tool, generates layout tool, renders preview tool — all automatically orchestrated.
• Leverage speech integration for on-site design: landscapers dictate observations while walking a property, and the AI processes them into design inputs.

What doesn't transfer:

• The SDK is infrastructure, not a product — it provides the building blocks but not the domain-specific logic, training data, or design intelligence needed for landscaping.
• The SDK's streaming patterns assume text-based output. Landscape design involves spatial data, images, and 3D models that require different streaming approaches.
• Tool calling patterns work well for discrete actions but may struggle with continuous spatial operations (dragging a plant across the canvas in real time).

Credit/pricing model:

• AI SDK: Free, open-source (MIT license), no usage costs
• Vercel AI Gateway (optional): Pay-as-you-go, zero markup on token costs; free tier included with Vercel account
• Model costs: Pass-through to the model provider (OpenAI, Anthropic, etc.) at their published rates
• Vercel hosting (if used): Standard Vercel pricing applies to the application, not the SDK
• The SDK itself has no credit system — the credit/pricing model is determined by the application builder (i.e., SimplyScapes would define its own credit model on top of the SDK)
• This is the only product in this analysis that is infrastructure rather than a consumer/prosumer product

---

Cross-Industry Insights

Patterns That Have Not Been Applied to Landscaping

1. Conversational-to-structured-output pipeline. Every product studied either generates flat images (Midjourney, ChatGPT, Firefly) or structured UI objects (Figma). No product in any market generates structured spatial designs through conversational input. The opportunity is to build a system where natural language ("I want a cozy fire pit area with native plantings") produces a structured landscape plan with named plants, dimensions, materials, and costs — not just a pretty picture.

2. Preference learning without explicit configuration. Midjourney's image-rating personalization and Canva's adaptive design suggestions represent passive preference capture. In landscaping, this could mean showing homeowners 20-30 landscape photos during onboarding, learning their aesthetic, and biasing all future suggestions — modern vs. traditional, formal vs. naturalistic, colorful vs. green.

3. Multi-turn editing with spatial awareness. ChatGPT's conversational image editing is powerful but spatially unaware. A landscaping-specific version could understand "move the pergola 3 feet closer to the house" as a precise spatial operation rather than an image manipulation.

4. Tiered quality/speed/cost generation. Runway's model tiering (Turbo vs. standard vs. Aleph) has not been applied to landscape design. Quick concept sketches, detailed plan views, and photorealistic client presentations are different products with different cost structures — they should be priced differently.

5. Direct-manipulation AI (non-prompt). Figma's erase/isolate/expand tools prove that AI does not require prompts. In landscaping, selecting a bare area and clicking "fill with seasonal color" is more intuitive than writing a prompt describing what should go there.

6. Enterprise credit pooling for design teams. Adobe's credit pool model and Figma's upcoming shared pool subscriptions address a real need: landscape design companies have 3-15 designers with variable AI usage. Per-user limits waste credits on light users and constrain heavy users.

7. Reference image consistency across views. Runway Gen-4's reference image system maintains character consistency across scenes. This pattern, applied to landscaping, would ensure that a design looks consistent whether shown as a plan view, a front elevation, or a 3D walkthrough.

8. Voice-driven on-site design. Midjourney's voice input + Draft Mode enables rapid ideation. For landscapers, this could enable on-site design sessions: walk the property, dictate observations and preferences into the app, and get real-time concept generation.

Architectural Insight

The Vercel AI SDK provides the exact infrastructure layer needed to implement most of these patterns in a Next.js application. Its provider abstraction means SimplyScapes can start with one model provider and switch or multi-route as the landscape AI model market develops. The tool-calling abstraction maps naturally to landscape design operations (add plant, check zone, calculate cost, generate rendering). The streaming architecture supports both conversational chat and progressive rendering.

---

Pricing Model Comparison

| Product | Model Type | Free Tier | Entry Paid | Mid Tier | Top Tier | Unit of Measurement | Overage Handling | Team/Pool Support | |---------|-----------|-----------|------------|----------|----------|---------------------|------------------|-------------------| | Canva AI | Bundled subscription | 50 total generations | $15/mo (500/mo) | $10/user/mo Teams (500/mo) | ~$20/user/mo Business | Per generation | Hard wall; must upgrade | Per-user only (no pooling) | | Figma AI | Credits per seat | 500/mo (Starter) | $5/editor/mo (3,000/mo) | $5/editor/mo Org (3,500/mo) | $5/editor/mo Enterprise (4,250/mo) | Credits (variable cost per action) | Enforcement begins Mar 2026 | Shared pool subscription (Mar 2026) | | Midjourney | GPU time subscription | None | $10/mo (3.3 GPU hrs) | $30/mo (15 GPU hrs + Relax) | $120/mo (60 GPU hrs + Relax) | GPU minutes | Relax Mode (unlimited, lower priority) | Per-account only | | Adobe Firefly | Credits per plan | Via free Creative Cloud | $9.99/mo (2,000 credits) | $29.99/mo (7,000 credits) | $199.99/mo (50,000 credits) | Credits (1 per image; 20-100/sec video) | On-demand credit purchase | Enterprise credit pool (optional add-on) | | Runway ML | Credits + unlimited tier | 125 credits (one-time) | $12/mo (625 credits) | $28/mo (2,250 credits) | $76/mo (unlimited Relax) | Credits ($0.01/credit; per-second for video) | Buy more or use Relax Mode | Per-organization | | ChatGPT | Rate-limited subscription | 2-3 images/day | $20/mo (50/3hrs) | — | $200/mo Pro (higher limits) | Images per time window | Wait for window reset | Per-user only | | Vercel AI SDK | Open-source (free) | Full SDK, no limits | N/A (pass-through to providers) | N/A | N/A | N/A (infrastructure layer) | N/A | N/A |

Key Pricing Takeaways for SimplyScapes

1. Credits are the dominant model — 5 of 7 products use some form of credit system (Canva, Figma, Adobe, Runway, and even Midjourney's GPU time is effectively credits). Time-window rate limiting (ChatGPT) is the exception, not the rule.

2. Variable credit cost by action complexity is best practice — Figma (30-100+ credits per action) and Adobe (1 credit/image, 20-100 credits/sec video) both price AI actions by complexity. A landscape design system should similarly charge differently for a quick concept sketch vs. a full photorealistic rendering.

3. "Unlimited at lower priority" reduces anxiety — Both Midjourney (Relax Mode) and Runway (Unlimited plan) offer unlimited generation at reduced priority. This is a powerful pattern for creative work where users need to experiment freely without watching a credit counter.

4. Free tiers drive adoption but must be tight — Canva (50 lifetime), ChatGPT (2-3/day), and Runway (125 one-time) all offer free access that is sufficient for evaluation but insufficient for real work. The ideal free tier for landscaping AI would let a homeowner generate 2-3 concept designs before requiring a paid plan.

5. Team credit pooling is an enterprise upsell — Adobe and Figma both offer or are introducing shared credit pools at premium rates. For landscape companies with variable designer utilization, this is a high-value feature worth charging for.

6. Bundling AI into existing subscriptions works for platforms with existing user bases — Canva and Adobe bundle AI credits into their existing plans. SimplyScapes, as a newer platform, may benefit more from a standalone AI credit model that can be priced and communicated clearly.

Patent Landscape & Freedom-to-Operate Assessment

Date: 2026-03-09 (Updated) Scope: AI-assisted landscape design, conversational image editing, credit-based pricing, intent routing, markup-guided generation, multi-turn dialogue for iterative image refinement Databases Searched: Google Patents, TDCommons Time Range: 2012-2026 CPC Codes: G06T 11/00, G06F 3/04845, G06N 3/08, G06F30/13, G10L15/1815, G06F17/2785 Search Date: 2026-03-09

---

1. Overall Risk Assessment: LOW-MODERATE

The combined feature set -- AI landscape design + conversational editing with object library + markup-guided generation + proactive clarification + credit billing -- represents a novel combination not directly claimed by any single patent or combination of patents. However, specific component technologies carry patent risk that must be managed through architectural differentiation and domain specificity.

Key Risk Summary:

• Adobe holds the strongest portfolio in conversational image editing (5+ granted US patents)
• Google/Alphabet has expanding coverage of diffusion-based text-guided editing techniques
• Home Outside, Inc. holds a recently granted patent specifically for AI landscape design generation
• Credit/token billing and markup-guided generation in the landscape context remain largely unencumbered
• No single patent covers the full SimplyScapes architecture (conversational + generative AI + landscape-specific + credit metering)

---

2. Patents by Topic Area

2.1 Conversational Image Editing (PRIMARY CONCERN)

| # | Patent | Title | Assignee | Filed | Status | Risk | |---|--------|-------|----------|-------|--------|------| | 1 | US11972757B2 (US20230148406A1) | Digital Media Environment for Conversational Image Editing and Enhancement | Adobe Inc. | 2023-01-03 (Priority 2018-08-22) | Granted, Active (exp. 2038) | HIGH | | 2 | US10579737B2 | Natural Language Image Editing Annotation Framework | Adobe Inc. | 2018-03-06 | Granted, Active (exp. 2038) | HIGH | | 3 | US11257491B2 | Voice Interaction for Image Editing | Adobe Inc. | 2018-11-29 | Granted, Active (exp. 2039) | HIGH | | 4 | US9412366B2 | Natural Language Image Spatial and Tonal Localization | Adobe Inc. | 2012-11-21 | Granted, Active (exp. 2034) | MODERATE | | 5 | US9436382B2 | Natural Language Image Editing | Adobe Inc. | 2012-11-21 | Granted, Active | MODERATE | | 6 | US20230230198A1 | Interactive Image Creation via NL Feedback (TiGAN) | Adobe Inc. | 2022 | Published | MODERATE | | 7 | EP4553759A2 | Image Editing Method, Apparatus, and Storage Medium (multi-round conversational) | Beijing Baidu Netcom | 2024-12-27 (Priority 2024-01-05) | Published | MODERATE | | 8 | WO2025209146A1 | Image Editing Method and Apparatus | ByteDance | 2025-03-13 (Priority 2024-04-01) | Published | LOW | | 9 | CN120655750A | Multi-round Image Modification Processing Method | Shenzhen Kukai Software | 2025-05-22 | Published | LOW | | 10 | US20200327884A1 | Customizable Speech Recognition System (creative applications) | Adobe Inc. | 2019-04-12 | Published | LOW |

Analysis: Adobe dominates this space with granted patents spanning from 2012 to 2024. The most concerning patent is US11972757B2, which describes a conversational digital image editing system with aesthetic scoring, intent mapping to canonical intentions, and iterative suggestion of editing operations. SimplyScapes should design its intent-routing architecture to differ from Adobe's "canonical intention mapping" approach.

Key Differentiator for SimplyScapes: Adobe's patents focus on photographic enhancement (exposure, contrast, color balance, aesthetic scores). SimplyScapes operates in landscape design (plant placement, hardscape, spatial layout, seasonal visualization) -- a fundamentally different editing domain. The object library concept (selecting from a curated catalog of landscape items and placing them via conversation) is a distinct, unclaimed combination.

2.2 Text-Guided Image Editing with Diffusion Models

| # | Patent | Title | Assignee | Filed | Status | Risk | |---|--------|-------|----------|-------|--------|------| | 1 | US20240037822A1 | Prompt-to-Prompt Image Editing with Cross-Attention Control | Google LLC | 2023-07-31 (Priority 2022-08-01) | Published | MODERATE | | 2 | JP7691548B2 | Text-Based Real-Life Image Editing Using Diffusion Models | Google LLC | 2024-04-18 (Priority 2023-04-18) | Granted | MODERATE | | 3 | CN119422137A | Hint-Driven Image Editing Using Machine Learning | Google LLC | 2024-05-09 (Priority 2023-05-09) | Published | MODERATE | | 4 | WO2024107884A1 | Null-Text Inversion for Editing Real Images Using Guided Diffusion Models | Google LLC | 2023-11-15 (Priority 2022-11-16) | Published | MODERATE | | 5 | EP4487299B1 | Fine-tuning Diffusion-Based Generative Neural Networks | Google LLC | 2024-03-13 (Priority 2023-03-17) | Granted | LOW | | 6 | US20240412458A1 | Diffusion-Guided Three-Dimensional Reconstruction | Google LLC | 2024-06-12 | Published | LOW | | 7 | US11983806B1 | Image Generation (inpainting/outpainting) | OpenAI | 2023 | Granted | MODERATE | | 8 | US11922550B1 | Hierarchical Text-Conditional Image Generation | OpenAI | 2023 | Granted | LOW | | 9 | US20220270310A1 | Web-Based Real-Time Image Editing with Neural Networks | Adobe Inc. | 2021 | Published | LOW |

Analysis: Google has 22+ patents covering diffusion-based image editing techniques. These are primarily implementation-specific (covering particular algorithms like cross-attention manipulation, null-text inversion) rather than broad application-layer claims. The risk is manageable if SimplyScapes uses third-party model APIs rather than implementing patented techniques directly.

Mitigation: Use third-party AI model APIs (OpenAI, Stability AI, etc.) for the generative layer. Model providers typically indemnify users against patent claims related to internal model techniques.

2.3 AI-Assisted Landscape / Design Generation

| # | Patent | Title | Assignee | Filed | Status | Risk | |---|--------|-------|----------|-------|--------|------| | 1 | US12518067B2 | System and Method for Generating a Landscape Design | Home Outside, Inc. | 2023-07-05 (Priority 2019-10-24) | Granted, Active (exp. 2041) | HIGH | | 2 | US20250225479A1 | Systems and Methods for 3D Model Visualization of Landscape Design | State Farm Mutual Auto Insurance | 2025-03-25 (Priority 2020-04-27) | Published | LOW | | 3 | CN118333571B | Intelligent Management System for Landscaping Engineering Projects | Liaocheng Zhengyuan | 2024-05-10 | Granted | LOW | | 4 | CN118940364B | Automatic Building Design Scene Generation Based on AI | Shenzhen Kuboo Architecture | 2024-07-17 | Granted | LOW | | 5 | US20210173968A1 | AI Systems for Interior Design | Realsee (Beijing) | 2021 | Published | LOW |

Analysis: The Home Outside patent (US12518067B2) is the most direct competitive concern. It describes a landscape design generation system with:

• A calculator engine, landscape design engine, and scoring engine
• Retrieving landscape data from online databases and comparing to existing designs
• Calculating landscape scores and generating improvements
• Displaying improved designs as 3D images

Key Differentiator for SimplyScapes: Home Outside's patent describes a score-based comparison system that retrieves data from online databases. SimplyScapes uses a conversational AI approach with generative diffusion models for visual generation, real-time iterative refinement, and domain-specific object libraries -- architecturally distinct from the patented system.

2.4 Intent Classification / Routing in Multimodal AI

| # | Patent | Title | Assignee | Filed | Status | Risk | |---|--------|-------|----------|-------|--------|------| | 1 | US9542949B2 | Multimodal Intent Satisfaction | Microsoft | Pre-2020 | Granted | LOW | | 2 | US12124508B2 | Multimodal Intent Discovery | Adobe Inc. | 2022 | Granted | LOW | | 3 | US9570070B2 | Multi-Modal Device Interactions in Voice Services | Amazon | Pre-2020 | Granted | LOW | | 4 | US11347801B2 | Multi-Modal Interaction with Automated Assistants | Google | Pre-2020 | Granted | LOW | | 5 | US10810274B2 | Optimizing Dialogue Policy Decisions for Digital Assistants | Apple Inc. | 2017-08-15 | Granted | LOW | | 6 | US10482874B2 | Hierarchical Belief States for Digital Assistants | Apple Inc. | 2017-08-15 | Granted | LOW | | 7 | US12197857B2 | Digital Assistant Handling of Personal Requests | Apple Inc. | 2021-07-15 | Granted | LOW |

Analysis: Existing intent classification patents are scoped to general-purpose voice/text assistants (Siri, Alexa, Google Assistant), not design-tool-specific routing between generation, editing, and clarification. SimplyScapes' intent routing between landscape-specific actions (add plant, modify hardscape, change season, refine area) is sufficiently differentiated.

2.5 Credit/Token-Based Pricing for AI Services

| # | Patent | Title | Assignee | Filed | Status | Risk | |---|--------|-------|----------|-------|--------|------| | 1 | US9197642B1 | Token-Based Billing for Rendering | Otoy | 2010 | Granted | LOW |

Analysis: No patents claim credit-based pricing specifically for AI image generation. The credit/token billing search returned 3,779 results across all generative AI billing topics, but none specifically cover the combination of credit metering for AI design generation services. Credit pricing is standard SaaS business practice with extensive prior art. SimplyScapes can implement credit/token-based billing without significant patent risk.

2.6 Markup/Annotation-Guided Generation

| # | Patent | Title | Assignee | Filed | Status | Risk | |---|--------|-------|----------|-------|--------|------| | 1 | US10825219B2 | Segmentation Guided Image Generation | Northeastern University | 2019 | Granted | LOW | | 2 | US12039431B1 | Multimodal ML Model Interaction (annotations) | OpenAI | 2023 | Granted | MODERATE | | 3 | WO2024158398A1 | AI Drawing Generation with Clarifying Questions | Individual | 2023 | Ceased | LOW |

Analysis: The search for markup-guided image generation patents returned mostly irrelevant results (medical imaging, video annotation, XAML troubleshooting). Using visual markup (user-drawn annotations, region selection) to guide AI generation does not have specific blocking patents in the landscape design context. OpenAI's US12039431B1 covers analysis of marked regions, not generation guided by canvas markup.

2.7 Multi-Turn Dialogue for Iterative Image Refinement

| # | Patent | Title | Assignee | Filed | Status | Risk | |---|--------|-------|----------|-------|--------|------| | 1 | US12142298B1 | Creating Digital Stories Based on Memory Graphs and Multi-Turn Dialogs | Meta Platforms, Inc. | 2023-01-09 (Priority 2022-02-14) | Granted | LOW | | 2 | JP6835398B2 | Multi-Turn Canned Dialog | Apple Inc. | 2018-10-29 | Granted | LOW | | 3 | EP4553759A2 | Image Editing Method (multi-round conversational) | Baidu | 2024-12-27 | Published | MODERATE |

Analysis: Multi-turn dialogue patents primarily cover general digital assistant functionality (Apple) or story generation (Meta). The Baidu patent on multi-round conversational image editing is the most relevant but primarily covers Chinese/EP jurisdictions. No active US patent broadly claims multi-turn conversational image editing with iterative refinement -- though Adobe's US11972757B2 partially covers this through its iterative suggestion mechanism.

---

3. Competitor IP Activity

Adobe Inc.

Portfolio Strength: VERY STRONG Adobe has the most comprehensive patent portfolio in NL/conversational image editing, with 5+ granted US patents spanning 2012-2024:

• Natural language spatial/tonal localization (foundational, 2012)
• Natural language image editing (2012)
• NL annotation framework for image editing (2018)
• Voice interaction for image editing (2018)
• Full conversational image editing environment (2018/2023, granted 2024)
• Customizable speech recognition for creative apps (2019)
• Interactive image creation with NL feedback (2022)
• Web-based real-time neural network editing (2021)

Adobe's Firefly product line commercially implements many of these patents. Their aggressive filing pattern suggests continued expansion.

Google/Alphabet

Portfolio Strength: STRONG (Technical Layer) Google has 22+ patents focused on underlying ML techniques:

• Prompt-to-prompt editing via cross-attention (2022)
• Null-text inversion for real image editing (2022)
• Text-based diffusion image editing (2023)
• Hint-driven image editing (2023)
• Diffusion model fine-tuning techniques (2023)
• 3D reconstruction and generation (2023)

Google's patents cover specific algorithms rather than application-layer products. Less likely to create broad blocking positions for vertical applications.

OpenAI

Portfolio Strength: LIMITED but GROWING

• Image generation including inpainting/outpainting (US11983806B1)
• Multimodal ML model interaction with annotations (US12039431B1)
• Hierarchical text-conditional image generation (US11922550B1)

OpenAI's DALL-E/ChatGPT products are significant commercially but their patent portfolio in this specific domain remains limited.

Canva

Portfolio Strength: NOT FOUND No relevant patents found. Canva's "Magic" AI features likely rely on third-party APIs and licensing.

Autodesk

Portfolio Strength: NOT FOUND (in AI image editing) Extensive traditional CAD/BIM portfolio but no filings in generative AI image editing or conversational design.

Home Outside, Inc.

Portfolio Strength: NICHE but DIRECT Single granted patent (US12518067B2) specifically covering AI landscape design generation with scoring. Most direct competitive threat in the landscape vertical, but describes a fundamentally different technical approach (database-comparison scoring) than generative AI chat interfaces.

Chinese Tech (Baidu, ByteDance)

Portfolio Strength: GROWING Both have recent filings covering multi-round conversational image editing. Primarily cover Chinese and European jurisdictions with limited direct US enforcement risk. Worth monitoring.

Stability AI / Midjourney

Portfolio Strength: NONE FOUND No patent applications identified. Primarily engaged in copyright litigation rather than patent prosecution.

---

4. Defensive Publication Findings (TDCommons)

Search of Technical Disclosure Commons for "AI image editing conversational" returned 22 defensive publications (2022-2026). Most relevant:

| # | Title | Authors | Date | Relevance | |---|-------|---------|------|-----------| | 1 | AI-driven Special Effects Generation and Application Framework for Visual Content | Colvin Pitts | 09/2025 | AI generation pipeline; establishes prior art for effect-based AI generation | | 2 | Assistive and Accessible User Interaction Mechanisms for AI-Powered Art | Alex Olwal, Shaun Kane | 09/2025 | Accessible interaction patterns for AI art; relevant to UI design | | 3 | Conversational Agent for Orchestrating Physical Fulfillment of Generative AI Outputs | Lance Nanek | 12/2025 | Conversational agent bridging AI outputs to physical actions; relevant to design-to-execution | | 4 | Generating Personalized Multimedia Content by Prompting Models with Location and Context Specific Inputs | Shiblee Hasan, Joseph Johnson Jr | 08/2025 | Location-specific generation; relevant to property-specific landscape generation | | 5 | Artificial Intelligence Based Creative Companion for Content Creation | Anonymous | 01/2022 | AI companion for creative workflows; prior art for AI creative assistants | | 6 | Policy-compliant Generative AI Deployment Using a Multimodal Critic | Xingyu Federico Xu et al. | 12/2025 | Safety/compliance for generative AI; relevant to content moderation | | 7 | Contextual Conversational Advertisements in Agents | Oded Elyada et al. | 12/2024 | Monetization of conversational AI; relevant to business model | | 8 | Three-Dimensional Mesh Editing Using Masked Large Reconstruction Models | Anonymous | 02/2025 | 3D editing with AI; relevant to future 3D landscape features | | 9 | A System and Method for Real-Time AI Driven Image Personalization for Digital Advertisements | Tarushi Dubey et al. | 02/2026 | Real-time AI image personalization; establishes prior art | | 10 | Generative Video Models to Create Panning Videos Anchored on a User Image | Lenord Melvix Joseph Stephen Max | 11/2023 | AI video from user images; establishes prior art for image-to-video in design context |

Key Takeaway: The defensive publication landscape is growing for conversational AI creative tools. Several publications establish prior art that could challenge overly broad patent claims. TDCommons coverage for design-specific AI tools remains thin, representing an opportunity for SimplyScapes to file defensive publications.

---

5. Freedom-to-Operate Assessment

5.1 High-Concern Areas

Adobe Conversational Image Editing Patents

• US11972757B2 broadly covers conversational image editing systems with aesthetic scoring and intent mapping
• Mitigation: Use LLM-based intent routing (function calling) rather than rule-based canonical intention mapping. Focus on landscape-specific intents rather than general photo editing. Avoid implementing aesthetic attribute scoring systems that mirror Adobe's claims.

Home Outside Landscape Design Patent

• US12518067B2 covers AI landscape design with calculator engine + scoring engine + database comparison
• Mitigation: Use generative AI (diffusion models) for visual generation rather than database-lookup design recommendation. Ensure the workflow is conversational/iterative, not score-and-improve. Do not implement a scoring engine that compares against online landscape databases.

5.2 Medium-Concern Areas

Google Diffusion Editing Techniques

• 22+ patents covering specific diffusion-based editing algorithms
• Mitigation: Use third-party model APIs where providers handle patent licensing. Avoid implementing specific patented techniques (cross-attention manipulation, null-text inversion) in custom code.

Baidu Multi-Round Conversational Editing

• EP4553759A2 covers multi-round conversational image editing
• Mitigation: Primarily Chinese/EP jurisdiction. Monitor for US continuation filings.

5.3 Low-Concern Areas

• Credit/token billing: No blocking patents; standard SaaS practice
• Markup/annotation-guided generation: No blocking patents in landscape context
• Intent routing for design tools: Existing patents scoped to general assistants, not vertical design tools
• Landscape-specific AI features: Plant libraries, climate-zone awareness, seasonal visualization, property context -- all largely unpatented

5.4 Recommended Actions

1. Architectural Differentiation: Use LLM-based function-calling for intent routing (not rule-based canonical intention mapping per Adobe's claims)

2. Domain Specificity: Frame patent-sensitive features within landscape design domain; "landscape design editing via conversation" is more defensible than "general image editing via conversation"

3. Model-Layer Separation: Use third-party AI model APIs for generation; model providers typically indemnify against technique-level patent claims

4. Design Around Home Outside: Use conversational AI + generative models rather than scoring-based database comparison systems

5. File Defensive Publications on TDCommons for:

   • Property-context-aware landscape generation from conversational AI
   • Climate zone and plant hardiness integration in generative AI design
   • Multi-turn landscape design refinement with spatial memory
   • Credit-based metering for domain-specific AI generation complexity
   • Object library integration with conversational generative editing

6. Patent Watch Program: Monitor quarterly filings from Adobe, Google, Baidu, ByteDance, and Home Outside in CPC classifications G06T11/00, G06F30/13, G10L15/1815

---

6. Search Methodology

6.1 Google Patents Searches Executed

1. "conversational image editing" -- 13 results
2. "multi-turn" "image editing" -- 37 results
3. "inpainting" "natural language" editing -- 584 results (top results sampled)
4. "markup guided" OR "annotation guided" image generation -- 15 results
5. "landscape design" AI generation -- 180 results (top results sampled)
6. "text guided" "image editing" "diffusion" assignee:Adobe -- searched
7. "natural language" "image editing" "intent" -- 1,664 results (top results sampled)
8. "image editing" "text prompt" "diffusion" assignee:Google -- 22 results
9. "generative fill" OR "text guided image editing" "diffusion model" -- searched
10. "credit system" OR "token based" "AI generation" OR "generative AI" billing -- 3,779 results (top sampled)
11. "image generation" "natural language" "user interface" "chat" assignee:Canva OR Autodesk OR OpenAI -- 0 results

6.2 TDCommons Search

• Query: AI image editing conversational -- 22 results across Defensive Publications Series

6.3 Company-Specific Patent Searches

• Adobe Inc. -- multiple targeted searches (strongest portfolio found)
• Google/Alphabet -- targeted diffusion editing search (22+ patents)
• OpenAI -- searched, limited relevant results
• Canva -- searched, no relevant results
• Autodesk -- searched, no relevant results in AI image editing
• Home Outside, Inc. -- found via landscape design search
• Baidu -- found via conversational image editing search
• ByteDance -- found via conversational image editing search
• Meta Platforms -- found via multi-turn dialog search
• Apple Inc. -- found via multi-turn dialog search (digital assistant focus)

6.4 Limitations

• Patent applications filed within the last 18 months may not yet be published
• Google Patents may not include all international jurisdictions
• This analysis does not constitute legal advice; formal FTO opinions should be obtained from patent counsel for specific product features
• Some companies may file under subsidiary or holding company names not captured in assignee searches

---

Patent findings are NOT legal advice. Consult qualified IP counsel for freedom-to-operate opinions before making product decisions based on this analysis.

Academic & Open Source Scan: Conversational AI Image Editing and Supporting Systems

Date: 2026-03-09 Scope: Academic papers, open source projects, and industry patterns relevant to building a conversational AI image editing interface for landscape/hardscape design. Sources: arXiv, Google Scholar, CVPR 2024, ECCV 2024, NeurIPS 2024/2025, ICLR 2024/2025, ACL/EMNLP, GitHub, industry reports.

---

1. Summary of Key Findings

Instruction-Based Image Editing has matured rapidly. The field has moved from single-instruction models (InstructPix2Pix) to MLLM-guided systems (SmartEdit, BrushEdit, GenArtist) that decompose complex edits into sub-tasks, self-verify results, and support multi-turn interaction. Agent-based architectures that orchestrate multiple specialized models via tool calling are now the dominant paradigm for production-quality editing.

Tool/Function Calling in LLMs is well-benchmarked (BFCL v4, ToolACE) and production-ready across major providers. The Vercel AI SDK (v6) and LangGraph provide mature TypeScript/Python frameworks for building tool-calling agents with structured outputs. Anthropic's Model Context Protocol (MCP) is emerging as a cross-platform standard for tool integration.

Model Routing can reduce LLM costs by 85%+ (RouteLLM) or match GPT-4 quality at 2% of cost (FrugalGPT). These techniques are directly applicable to routing simple edits to lightweight models while reserving expensive models for complex reasoning.

Multi-Turn Visual Dialogue systems like DialogGen and Talk2Image demonstrate that multi-agent architectures outperform single-agent approaches for complex editing workflows, addressing "intention drift" in long conversations.

Generative UI is an emerging pattern where LLMs dynamically render custom interface components (Vercel AI SDK, Google A2UI, assistant-ui). This is a natural fit for a chat-based image editing interface that needs to present different controls depending on the editing context.

Usage-Based Billing for AI services has converged on credit-based systems. Stripe's $1B acquisition of Metronome (Dec 2025) validates the market. Open source options (Lago, FlexPrice) exist for self-hosted billing infrastructure.

---

2. Reference Table

| # | Type | Reference | Venue/Date | Relevance | |---|------|-----------|------------|-----------| | Instruction-Based Image Editing | | | | | | 1 | Paper | InstructPix2Pix: Learning to Follow Image Editing Instructions | CVPR 2023 | Foundation model for text-instruction image editing; baseline for all subsequent work | | 2 | Paper | SmartEdit: Complex Instruction-based Image Editing with MLLMs | CVPR 2024 | Uses MLLM for complex instruction understanding; Bidirectional Interaction Module; Reason-Edit benchmark | | 3 | Paper | BrushNet: Plug-and-Play Image Inpainting with Dual-Branch Diffusion | ECCV 2024 | Plug-and-play inpainting architecture; decomposed dual-branch diffusion; BrushData/BrushBench | | 4 | Paper | BrushEdit: All-In-One Image Inpainting and Editing | arXiv 2412.10316 (2024) | Agent-cooperative framework with MLLM + inpainting; free-form instruction editing; category classification + mask acquisition pipeline | | 5 | Paper | MGIE: Guiding Instruction-based Image Editing via MLLMs | ICLR 2024 | Apple's MLLM-guided editing; derives expressive instructions from ambiguous user input; open source | | 6 | Paper | Emu Edit: Precise Image Editing via Recognition and Generation Tasks | CVPR 2024 | Meta's multi-task editing model; 7-task benchmark; task embeddings for generalization | | 7 | Paper | GenArtist: Multimodal LLM as Agent for Unified Image Generation and Editing | NeurIPS 2024 (Spotlight) | MLLM agent with tool library; tree-structured planning; self-correction with verification; 7%+ improvement over DALL-E 3 | | 8 | Paper | InstructDiffusion: A Generalist Modeling Interface for Vision Tasks | CVPR 2024 | Unified framework casting vision tasks as pixel manipulation via human instructions | | 9 | Paper | OmniGen: Unified Image Generation | arXiv 2409.11340 (2024) | Simplified architecture; no ControlNet/IP-Adapter needed; unified transformer for text+image | | 10 | Paper | UltraEdit: Instruction-based Fine-Grained Image Editing at Scale | NeurIPS 2024 | 4M editing sample dataset; region-based editing; anchored on real images | | 11 | Paper | AnyEdit: Mastering Unified High-Quality Image Editing | arXiv 2411.15738 (2024) | 2.5M editing pairs across 20+ edit types; task-aware routing; AnyEdit-Test benchmark | | 12 | Paper | ImgEdit: A Unified Image Editing Dataset and Benchmark | NeurIPS 2025 | 1.2M edit pairs; multi-turn editing tasks; comprehensive benchmark | | 13 | Paper | Step1X-Edit: Open Source Image Editing Framework | 2025 | MLLM-based instruction processing; open source; ComfyUI integration | | Tool Calling / Function Calling | | | | | | 14 | Paper | Gorilla: Large Language Model Connected with Massive APIs | NeurIPS 2024 | Retriever-Aware Training (RAT); APIBench; surpasses GPT-4 on API calls | | 15 | Paper | ToolACE: Winning the Points of LLM Function Calling | ICLR 2025 | Self-evolution synthesis for tool-learning data; 26,507 API pool; SOTA on BFCL | | 16 | Benchmark | Berkeley Function Calling Leaderboard (BFCL) v4 | Ongoing (2024-2025) | Industry-standard benchmark; AST evaluation; v3 adds multi-turn; v4 adds agentic evaluation | | 17 | Paper | Octopus v2: On-device Language Model for Super Agent | arXiv 2024 | Functional token strategy; 99.5% accuracy; 140x faster inference than RAG; 2B parameters | | Model Routing & Cascading | | | | | | 18 | Paper | RouteLLM: Learning to Route LLMs with Preference Data | ICLR 2025 | Routes between strong/weak LLMs; 85% cost reduction on MT Bench; 45% on MMLU; open source framework | | 19 | Paper | FrugalGPT: How to Use LLMs While Reducing Cost | arXiv 2305.05176 (2023) | LLM cascade approach; matches GPT-4 at 2% cost; 4% accuracy improvement at same cost | | 20 | Paper | A Unified Approach to Routing and Cascading for LLMs | ETH Zurich 2024 | Theoretical unification of routing and cascading strategies | | Multi-Turn Visual Dialogue | | | | | | 21 | Paper | DialogGen: Multi-modal Interactive Dialogue System for Multi-turn T2I Generation | arXiv 2403.08857 (2024) | MLLM + T2I integration; multi-turn generation quality enhancement | | 22 | Paper | Talk2Image: Multi-Agent System for Multi-Turn Image Generation and Editing | arXiv 2508.06916 (2025) | Multi-agent architecture; addresses intention drift and incoherent edits | | 23 | Paper | DialogPaint: A Dialog-based Image Editing Model | arXiv 2303.10073 (2023) | Conversational image editing through natural dialogue; iterative multi-round editing | | 24 | Paper | TDRI: Two-Phase Dialogue Refinement for Interactive Image Generation | arXiv 2503.17669 (2025) | Initial Generation Phase + Interactive Refinement Phase; handles ambiguous prompts | | Multimodal UI/UX Design Patterns | | | | | | 25 | Paper | MAUI: Multimodal AI-augmented UI Development Architecture | Stanford 2024 | Explicit instruction + implicit preference handling; ReactGenie and AMMA frameworks | | 26 | Paper | A Multimodal GUI Architecture for LLM-Based Conversational Assistants | arXiv 2510.06223 (2025) | Strong cohesion between GUI and linguistic UI | | 27 | Paper | Generative Interfaces for Language Models | arXiv 2508.19227 (2025) | Theoretical framework for LLM-generated interfaces | | 28 | Report | Generative UI Report 2025 | Thesys 2025 | Industry survey of generative UI patterns and adoption | | Credit/Billing Systems | | | | | | 29 | Report | AI Pricing in Practice: 2025 Field Report | Metronome 2025 | Cost-plus credit systems (30-50% markup); customer anxiety about unpredictable costs | | 30 | Report | Token-Based Pricing: How to Account for AI Credits | Afternoon 2024 | Accounting and revenue recognition for credit-based AI billing | | Protocols & Standards | | | | | | 31 | Standard | Model Context Protocol (MCP) | Anthropic Nov 2024 | Open standard for tool integration; adopted by OpenAI, Google; SDKs in TS/Python/C#/Java |

---

3. Detailed Analysis by Topic Area

3.1 Instruction-Based Image Editing

The evolution of instruction-based image editing follows a clear trajectory from simple to complex:

Generation 1 - Direct Instruction (2022-2023): InstructPix2Pix established the paradigm of editing images from natural language instructions. It uses paired training data generated by combining GPT-3 (for instruction generation) and Stable Diffusion (for image pairs). The limitation is reliance on the CLIP text encoder, which struggles with complex or ambiguous instructions.

Generation 2 - MLLM-Guided Editing (2024): SmartEdit (CVPR 2024) introduced the use of Multimodal Large Language Models to understand complex editing instructions. Its Bidirectional Interaction Module enables comprehensive information flow between the input image and MLLM output. The Reason-Edit benchmark specifically targets complex instruction scenarios where prior methods fail.

MGIE (ICLR 2024, Apple) takes a similar approach but focuses on deriving "expressive instructions" from ambiguous user input. For example, it converts vague instructions like "make the sky more blue" into precise operations like "increase sky saturation by 20%." This translation step is critical for production systems where users give imprecise instructions.

Emu Edit (CVPR 2024, Meta) introduced multi-task training across editing types plus computer vision tasks (segmentation, keypoint detection). Its task embedding approach enables rapid adaptation to new editing tasks with minimal examples, which is valuable for extending the system to domain-specific edits (e.g., landscape-specific operations).

Generation 3 - Agent-Based Editing (2024-2025): GenArtist (NeurIPS 2024 Spotlight) represents the current state of the art. It uses an MLLM as an agent that:

1. Decomposes complex requests into sub-tasks
2. Selects appropriate tools from a library of specialized models
3. Constructs a planning tree for execution
4. Verifies results at each step and self-corrects

This architecture achieves 7%+ improvement over DALL-E 3 on T2I-CompBench and state-of-the-art on MagicBrush. The agent pattern maps directly to a chat interface where the LLM orchestrates multiple image processing operations.

BrushEdit extends this with a specific focus on inpainting workflows. Its agent pipeline performs: editing category classification, main object identification, mask acquisition, and editing area inpainting. The dual-branch architecture (from BrushNet, ECCV 2024) handles arbitrary mask shapes without needing separate models for different mask types.

Datasets and Benchmarks: The field has produced increasingly large and diverse training datasets:

• UltraEdit (NeurIPS 2024): 4M editing samples, region-based, anchored on real images
• AnyEdit (2024): 2.5M pairs across 20+ edit types with task-aware routing
• ImgEdit (NeurIPS 2025): 1.2M pairs including multi-turn editing tasks
• BrushData/BrushBench: Segmentation-based inpainting benchmarks

Key Takeaway for SimplyScapes: The agent-based approach (GenArtist/BrushEdit pattern) is the most promising architecture. An MLLM orchestrator can decompose landscape editing instructions ("add a stone patio behind the garden beds") into sub-operations (identify garden area, determine patio placement, generate stone texture, blend into scene) using specialized tools. The translation of ambiguous instructions into precise operations (MGIE pattern) is essential for consumer-facing products.

---

3.2 Tool Calling / Function Calling in LLMs

Benchmarking and Evaluation: The Berkeley Function Calling Leaderboard (BFCL) has become the industry standard for evaluating LLM tool-calling capabilities. Now at version 4, it has evolved through:

• v1: AST-based evaluation metric for function call correctness
• v2: Enterprise and community-contributed function definitions
• v3: Multi-turn interactions (critical for conversational editing)
• v4: Holistic agentic evaluation (complete task workflows)

ToolACE (ICLR 2025) demonstrated that purpose-built training data for tool calling can enable an 8B parameter model to rival GPT-4's function calling performance. Its self-evolution synthesis process generated 26,507 diverse APIs for training, suggesting that domain-specific tool-calling fine-tuning (e.g., for image editing APIs) is feasible.

On-Device Function Calling: Octopus v2 (Nexa AI, 2024) achieved 99.5% accuracy in function-calling tasks at 2B parameters using a "functional token" strategy that reduces context length by 95%. This approach could enable lightweight, on-device routing of simple editing operations without cloud API calls.

Production Frameworks: Three frameworks dominate the tool-calling implementation space:

1. Vercel AI SDK (v6): Unifies structured output generation with tool calling. The Agent abstraction enables reusable agent definitions with model, instructions, and tools. Supports streaming React Server Components for generative UI. TypeScript-native.

2. LangGraph (LangChain): Recommended for production agents as of 2025. Graph-based state machine for multi-agent workflows. Supports MCP, streamable HTTP, and OpenAPI-based tool calls. Better suited for complex stateful workflows with branching logic.

3. Anthropic Claude Tool Use: First-class structured outputs with strict: true guaranteeing schema validation. Agent SDK supports JSON Schema, Zod, or Pydantic for output validation. MCP for cross-platform tool integration.

Model Context Protocol (MCP): Announced by Anthropic in November 2024, MCP is an open standard for connecting AI assistants to external tools and data sources. Adopted by OpenAI and Google DeepMind. SDKs available in TypeScript, Python, C#, and Java. Pre-built servers exist for GitHub, Slack, Postgres, and other enterprise systems. Security concerns were raised in April 2025 regarding prompt injection and tool permission vulnerabilities.

Key Takeaway for SimplyScapes: The Vercel AI SDK is the natural choice given the Next.js stack. Its tool-calling + structured output unification in v6 and generative UI support via React Server Components directly address the need for a chat interface that renders custom editing controls. MCP provides future-proofing for tool integration, though security hardening is needed.

---

3.3 Model Routing and Cascading

RouteLLM (ICLR 2025, LMSYS): Proposes router models that dynamically select between a stronger and weaker LLM per query based on estimated difficulty. Key results:

• 85% cost reduction on MT Bench vs. GPT-4-only
• 45% cost reduction on MMLU
• 35% cost reduction on GSM8K The routing approach selects a single model per query (not cascading), which keeps latency constant. The framework is open source.

FrugalGPT (Stanford, 2023): Uses an LLM cascade approach: sequentially query models from cheapest to most expensive until a reliable response is obtained. A learned scoring function decides when to accept a response. Key results:

• Match GPT-4 quality at 2% of cost
• Exceed GPT-4 accuracy by 4% at the same cost Trade-off: latency increases with cascade depth.

ETH Zurich Unified Approach (2024): Provides a theoretical framework unifying routing (single model selection) and cascading (sequential model queries) under a common optimization objective, enabling hybrid strategies.

Key Takeaway for SimplyScapes: For a conversational image editing system, a hybrid routing strategy is optimal:

• Simple operations (crop, rotate, brightness): Route to lightweight models or deterministic functions (no LLM needed)
• Standard edits (object removal, style transfer): Route to mid-tier models (e.g., fine-tuned Stable Diffusion)
• Complex creative edits (scene composition, multi-step transformations): Route to full MLLM agent with tool calling This tiered approach can dramatically reduce per-edit costs while maintaining quality where it matters.

---

3.4 Multi-Turn Visual Dialogue Systems

Core Challenge - Intention Drift: Talk2Image (2025) identifies the key problem in multi-turn image editing: single-agent systems suffer from "intention drift" where cumulative user goals become misaligned as the conversation progresses, leading to incoherent edits. Multi-agent architectures address this by separating concerns (understanding, planning, execution, verification).

DialogGen (2024): Equips MLLMs with text-to-image models to extend output modality. Focuses on maintaining generation quality across multiple conversation turns through strong multi-modal comprehension. Key innovation: the MLLM maintains conversation context while delegating generation to specialized models.

TDRI (2025): Introduces a two-phase approach:

1. Initial Generation Phase: Interpret the user's first request and generate a baseline
2. Interactive Refinement Phase: Iteratively refine based on follow-up instructions

This maps well to a landscape editing workflow where users start with a general request and progressively refine specific areas.

DialogPaint (2023): Early demonstration of multi-round image editing through natural dialogue. Establishes the pattern of maintaining edit history for undo/redo within the conversation flow.

Key Takeaway for SimplyScapes: Multi-turn editing requires explicit state management to prevent intention drift. The two-phase approach (generate-then-refine) from TDRI is a practical pattern. A multi-agent architecture (Talk2Image) should be considered for complex editing flows, with separate agents for understanding intent, selecting tools, executing edits, and verifying quality.

---

3.5 Multimodal LLM UI/UX Design Patterns

Generative UI: The most significant UX trend is generative UI, where the LLM dynamically renders custom interface components based on conversation context. Key implementations:

• Vercel AI SDK: Streams React Server Components from the server, allowing the LLM to render custom UI elements (sliders, image previews, comparison views) within the chat flow.
• Google A2UI (2025): An open project for agent-driven interfaces. Uses a flat list of components with ID references that LLMs can generate incrementally, enabling progressive rendering. Supports incremental UI updates based on conversation progression.
• assistant-ui: Open source TypeScript/React library for AI chat with tool call rendering, human approval workflows, and safe frontend actions.

Architectural Patterns: The Stanford MAUI architecture identifies two interaction modes:

1. Explicit direct instructions: User tells the system what to do ("remove the tree")
2. Implicit preference inference: System learns from user behavior and feedback

Most current interfaces fall into the "chatbot on the side" category, but the trend is toward strong cohesion between GUI elements and the conversational interface.

Post-Chat UI (2025): Industry analysis identifies the shift beyond pure chat interfaces. LLMs can generate context-specific controls, buttons, layouts, and navigation tailored to the current task. This is particularly relevant for image editing where a chat-only interface is insufficient -- users need visual controls for spatial operations.

Key Takeaway for SimplyScapes: The interface should blend conversational AI with generated UI components. When a user asks to adjust brightness, the system should render a slider within the chat flow (generative UI). When they ask to select a region, it should render an interactive canvas overlay. The Vercel AI SDK's React Server Component streaming and the A2UI incremental rendering pattern are directly applicable.

---

3.6 Credit/Token Billing Systems for AI Services

Industry Trends: Over 61% of new B2B SaaS products are exploring usage-based pricing (OpenView Partners, 2024). Stripe's $1B acquisition of Metronome in December 2025 validates the credit-based billing infrastructure market.

Common Architecture: The dominant pattern is a cost-plus credit system:

1. Credits are allocated (purchased or included in subscription tier)
2. Credits are consumed based on operation complexity (input/output tokens, GPU time, model tier)
3. Standard markup of 30-50% over raw API costs
4. Real-time dashboards for usage visibility
5. Usage alerts to prevent bill shock

Key Challenge - Customer Anxiety: Multiple industry reports flag unpredictable costs as the top barrier to adoption. Buyers do not understand credit burn rates. Solutions include:

• Real-time usage dashboards
• Spend alerts and budget caps
• Predictable credit packages (e.g., "100 edits per month")
• Value-based credit pricing (charge per edit, not per token)

Infrastructure Options:

• Stripe Billing + Metronome: Enterprise-grade, now integrated. Handles 100K+ events/sec.
• Lago (open source): Self-hosted billing with event-driven metering. $34M funding. GitHub: github.com/getlago/lago
• FlexPrice (open source): AI-native billing with credit/top-up support. Self-hosted. GitHub: github.com/flexprice/flexprice
• Amberflo: Real-time usage metering focused on cost allocation across AI models
• Orb: Usage-based billing for AI companies (proprietary)

Key Takeaway for SimplyScapes: A credit-based system mapped to user-understandable units ("design edits" rather than "tokens") reduces customer anxiety. The tiered model routing strategy (Section 3.3) directly feeds into credit pricing: simple edits cost 1 credit, complex edits cost 5-10 credits. Lago or FlexPrice can provide the billing infrastructure. Stripe's native integration with Metronome simplifies payment processing.

---

4. Open Source Projects

4.1 Image Editing Models & Frameworks

| Project | GitHub | Stars | Description | |---------|--------|-------|-------------| | ComfyUI | comfyanonymous/ComfyUI | 89K+ | Node-based UI for Stable Diffusion; extensible with custom nodes; supports Flux, SDXL, SD3 | | AUTOMATIC1111 WebUI | AUTOMATIC1111/stable-diffusion-webui | 140K+ | Most popular SD web UI; extensive extension ecosystem | | BrushNet | TencentARC/BrushNet | - | Plug-and-play inpainting; ECCV 2024 | | BrushEdit | TencentARC/BrushEdit | - | Agent-cooperative inpainting + editing pipeline | | MGIE (Apple) | apple/ml-mgie | - | MLLM-guided image editing; ICLR 2024 | | GenArtist | zhenyuw16/GenArtist | - | MLLM agent for unified generation + editing; NeurIPS 2024 Spotlight | | InstructPix2Pix | timothybrooks/instruct-pix2pix | - | Foundation instruction-based editing model | | InstructDiffusion | cientgu/InstructDiffusion | - | Generalist vision task interface; CVPR 2024 | | OmniGen | VectorSpaceLab/OmniGen | - | Unified image generation without extra modules | | UltraEdit | pkunlp-icler/UltraEdit | - | 4M editing sample dataset + models; NeurIPS 2024 | | Step1X-Edit | stepfun-ai/Step1X-Edit | - | Open source MLLM-based image editing; 2025 | | Qwen-Image | QwenLM/Qwen-Image | - | Foundation model for image generation + editing |

4.2 Tool Calling & Agent Frameworks

| Project | GitHub | Description | |---------|--------|-------------| | Gorilla | ShishirPatil/gorilla | LLM for API calls; Berkeley Function Calling Leaderboard; NeurIPS 2024 | | ToolACE-8B | HuggingFace: Team-ACE/ToolACE-8B | SOTA function calling at 8B parameters; ICLR 2025 | | Vercel AI SDK | vercel/ai | TypeScript SDK for AI apps; tool calling + generative UI; v6 with Agent abstraction | | LangChain / LangGraph | langchain-ai/langchain | Python/TS agent framework; graph-based workflows; MCP support | | MCP (Model Context Protocol) | modelcontextprotocol | Anthropic's open standard for tool integration; TS/Python/C#/Java SDKs | | RouteLLM | lm-sys/RouteLLM | Open source LLM routing framework; ICLR 2025 |

4.3 UI Frameworks for AI Chat

| Project | GitHub | Description | |---------|--------|-------------| | assistant-ui | assistant-ui/assistant-ui | TypeScript/React library for AI chat; tool call rendering; human approval flows | | Vercel AI Chatbot Template | vercel/ai-chatbot | Next.js + AI SDK reference implementation with generative UI | | Google A2UI | Announced 2025 | Open project for agent-driven interfaces; flat component list with incremental updates |

4.4 Billing Infrastructure

| Project | GitHub | Description | |---------|--------|-------------| | Lago | getlago/lago | Open source metering + usage-based billing; event-driven; $34M funding | | FlexPrice | flexprice/flexprice | Open source usage-based billing; credits + top-ups; AI-native focus |

---

5. Key Takeaways for SimplyScapes Implementation

5.1 Architecture Recommendation: Agent-Based Editing with Tool Calling

The research strongly supports an architecture where:

• A conversational MLLM (e.g., Claude, GPT-4) serves as the orchestrator
• It calls specialized image editing tools via function calling (Vercel AI SDK tool definitions)
• A planning layer decomposes complex editing requests into sub-tasks (GenArtist pattern)
• A verification step checks results and triggers self-correction if needed
• State management tracks conversation history and edit history to prevent intention drift

5.2 Model Routing for Cost Optimization

Implement a tiered routing strategy: | Tier | Operations | Model | Est. Cost | |------|-----------|-------|-----------| | 0 - Deterministic | Crop, rotate, resize, brightness/contrast | No LLM needed (Sharp/Canvas API) | Near zero | | 1 - Lightweight | Simple object removal, background blur, color adjustment | Fine-tuned SD model or API | Low | | 2 - Standard | Style transfer, material swap, object addition | Mid-tier image model (SDXL/Flux) | Medium | | 3 - Complex | Scene composition, multi-element design, creative interpretation | Full MLLM agent + specialized tools | High |

A RouteLLM-style classifier at the conversation level determines which tier handles each request.

5.3 UI Pattern: Hybrid Chat + Generative UI

Based on the research, the optimal interface combines:

1. Chat thread for natural language interaction and edit history
2. Generative UI components rendered within the chat (sliders, color pickers, region selectors) via Vercel AI SDK RSC streaming
3. Persistent canvas for image display with interactive overlays for spatial operations
4. Before/after comparison views generated dynamically based on edit type
5. Approval gates for destructive or high-credit operations (assistant-ui pattern)

5.4 Credit System Design

Map billing to user-understandable units:

• Simple edits (Tier 0-1): 1 credit
• Standard edits (Tier 2): 3-5 credits
• Complex edits (Tier 3): 8-15 credits
• Batch operations: Volume discount
• Implement real-time credit counter in the UI
• Offer credit packages aligned to common use cases (e.g., "Design Preview Pack: 50 edits")
• Use Lago or FlexPrice for backend metering; Stripe for payment processing

5.5 Multi-Turn Conversation Design

Critical patterns from the research:

1. Maintain explicit edit state -- track all edits as a stack for undo/redo
2. Two-phase workflow (TDRI): Generate initial design, then iterative refinement
3. Intention summarization: Periodically summarize the user's cumulative intent to prevent drift
4. Context windowing: For long sessions, compress early conversation turns while preserving key decisions
5. Multi-agent separation: Separate understanding, planning, execution, and verification into distinct agent roles (even if running on the same model)

5.6 Competitive Advantage Opportunities

Based on the gap analysis across all research:

• No existing system combines landscape-specific editing with conversational AI -- the domain specialization is the moat
• Agent-based architectures are published but not yet productized for consumer verticals -- there is a first-mover window
• Cost optimization via routing is well-researched but rarely implemented in consumer image editing products
• Generative UI for image editing is nascent -- most tools still use traditional UI with optional AI features
• Multi-turn editing with state management is an unsolved UX problem at scale -- the research identifies the problems but production solutions are scarce

---

6. Research Gaps and Future Directions

1. Domain-specific editing evaluation: No benchmarks exist for landscape/hardscape editing quality. Creating a domain-specific evaluation set would enable systematic quality improvement.

2. Cost-quality Pareto frontier: While RouteLLM and FrugalGPT optimize for general LLM tasks, no published work optimizes the cost-quality trade-off specifically for image editing pipelines with heterogeneous models.

3. Real-time collaborative editing: All multi-turn dialogue research assumes a single user. Collaborative design review (homeowner + designer) in a shared editing session is unexplored.

4. Deterministic + generative hybrid pipelines: Most research focuses purely on generative approaches. Combining deterministic rendering (3D landscape visualization) with generative editing (style/material transfer) is a gap.

5. User intent disambiguation for spatial operations: How to efficiently resolve ambiguous spatial references ("the area near the fence") in landscape images is under-explored in the academic literature.

Professional Landscape Designer & Architect Workflow

Date: 2026-05-01 | Purpose: Establish the canonical phases of a professional landscape design project and identify which phases map to AI-assisted edits on a rendered image. Audience: UX Designer, Product Lead — input for the Visual Designer AI Tab guided workflow.

TL;DR

Professional landscape design follows a five-phase progression — Site Analysis → Programming → Schematic Design → Design Development → Construction Documentation — codified by the American Society of Landscape Architects (ASLA) and taught at every accredited landscape architecture program. Most homeowner-facing AI design tools collapse this into a single one-shot transform, which leaves the entire design development phase (where most decisions actually happen) on the table. The phases that map cleanly to AI-assisted edits on a rendered image are 2 (Programming → bubble-diagram-equivalent), 3 (Schematic → "reduce grass to a central shape"), and 4a (early Design Development → "add foundation shrubs and bedding plants"). Site Analysis (Phase 1) and full Construction Documentation (Phase 5) require dimensional, measured drawings outside the scope of an image-edit workflow.

---

1. The Canonical Five Phases (ASLA Standard)

The American Society of Landscape Architects defines five phases in their Landscape Architecture Documentation Standards, taught in accredited programs and used by professional firms worldwide. Industry guides, university extension programs, and design textbooks all use variations of the same five-phase model.

| # | Phase | Other Names | Primary Question Answered | |---|-----------------------------|--------------------------|-----------------------------------------------| | 1 | Site Analysis / Pre-Design | Site Inventory, Survey | "What is here, and what does it constrain?" | | 2 | Programming | Needs Assessment, Brief | "What functions must the design serve?" | | 3 | Schematic Design | Conceptual, Concept Plan | "What is the spatial organization?" | | 4 | Design Development | Preliminary, Master Plan | "What does it actually look like?" | | 5 | Construction Documentation | CDs, Construction Set | "How does it get built?" |

Sources: ASLA Documentation Standards, NC State Extension — Landscape Design, University of Florida IFAS — Landscape Design Principles, Verdance Landscape Architecture — Preliminary Design Phase, Loft Six Four — A Walk Through the Landscape Architecture Process.

Phase 1: Site Analysis (Pre-Design)

What happens. The designer documents what exists on the site and its surrounding context. This is observation-and-inventory work, not creative design.

Inputs. Property survey, photographs, soil tests, on-site walks, solar/shade studies, local zoning and codes.

Activities.

• Locate and measure existing site features — trees, structures, utilities, slopes, drainage patterns
• Document environmental factors — solar orientation, prevailing winds, microclimate, soil type, water table
• Note context — adjacent properties, sightlines, sounds, neighborhood character
• Identify constraints — easements, setbacks, code restrictions, protected vegetation

Artifacts produced.

• Base plan: A scaled drawing (typically 1"=10' for residential) showing property lines, existing structures, and important features
• Site inventory overlay: Tracing-paper overlays documenting solar patterns, drainage, sightlines, problem zones
• Site analysis diagram: A summary graphic identifying assets ("specimen oak — preserve") and liabilities ("poor drainage — divert")

Duration. 1–4 weeks depending on site complexity.

AI workflow mapping. Site analysis is fundamentally a measured, dimensional activity — it requires a scaled base plan, not a photograph. Site analysis is therefore out of scope for image-edit workflows. However, the SimplyScapes platform already addresses this elsewhere (Site Planner, Property Onboarding) and produces a property model that can supply context to the AI tab. The AI tab can consume site-analysis outputs (sun exposure tags on regions of the image, climate zone) without performing site analysis itself.

Phase 2: Programming

What happens. The designer translates the client's needs and desires into a written program that lists required spaces, functions, and relationships. This is a brief — not a design.

Inputs. Client interview, lifestyle questionnaire, family size, entertaining habits, accessibility needs, budget.

Activities.

• Interview the client about how they will use the space
• Categorize uses (public/social, private/family, utility/service)
• Identify required spaces ("outdoor kitchen," "kids' play area," "vegetable garden," "screening from neighbor")
• Establish priorities (must-have vs. nice-to-have)
• Set the maintenance threshold (high/medium/low)

Artifacts produced.

• Program list: Written list of required functions and spaces
• Adjacency matrix: Which spaces should be near each other; which should be separated
• Bubble diagram (preliminary): First spatial sketch showing function zones as bubbles. NC State Extension calls this "the most important step the designer ever takes" because it locks in organization before any aesthetics.

Duration. 1–2 weeks. Often happens in a single client meeting plus follow-up.

AI workflow mapping. Programming is the closest analogue to "asking the user what they want." A click-driven questionnaire that captures functional needs (entertaining? kids? vegetables? privacy?) and spatial preferences (formal? wild? low-maintenance?) maps almost 1:1 to the programming phase. This is the right place for the AI tab's lead-in questionnaire.

Phase 3: Schematic Design

What happens. The designer translates the program into spatial form — where each function lives on the site, in what shape, with what circulation between them. Aesthetics begin to emerge but material specifics are deferred.

Inputs. Site analysis + program list + bubble diagram + style references (inspiration photos, mood boards).

Activities.

• Refine bubble diagrams into shaped spaces with edges and proportions
• Test multiple spatial alternatives ("Concept A, B, C")
• Establish circulation paths and view corridors
• Lock in major hardscape footprints (patio location, walk paths, pool position)
• Choose a design vocabulary — geometric/formal vs. organic/naturalistic

Artifacts produced.

• Concept plan(s): Multiple alternative layouts, often presented as 2–3 options for the client to choose between
• Schematic master plan: The chosen concept refined into a single scaled plan view
• Loose elevation sketches: Quick studies of key views

Duration. 2–4 weeks.

AI workflow mapping. Schematic Design is the strongest fit for AI-assisted image edits in the early flow. This is exactly where "reduce the grass to a central shape," "carve a planting bed out of the lawn," "place a patio in this corner" — the Define Spaces and Layout step in the SimplyScapes brief — belongs. The user is shaping spatial organization, not yet picking specific plants or materials.

Phase 4: Design Development

What happens. The schematic becomes a real design. Specific materials are chosen, plant species are named, lighting is positioned, custom features are detailed. This is "the most artful rendition" the designer produces — a fully visualized design with realistic representation of plants and materials. (Verdance calls Preliminary/Design Development "the purest articulation of the design concept.")

Inputs. Approved schematic + material samples + plant catalog + budget.

Activities (in order).

4a. Base Plantings. Trees first (largest, longest-lived, set bones of design). Then foundation shrubs. Then perennials and groundcovers. The professional rule of thumb: 5–10 plant varieties selected and mass-planted repetitively rather than 30 specimens scattered. Most firms use 20–30 core plants property-wide with 10–15 accents (per The Landscape Library and University of Florida MG086).

4b. Hardscape Materials. Specific stone, paver, or wood selections. Limit to 2–3 primary hardscape materials repeated for cohesion.

4c. Decor & Finishing Touches. Lighting, water features, sculptures, planters, furniture. The "jewelry" that completes the composition.

4d. Detail Drawings. Section cuts, custom feature details, specifications.

Artifacts produced.

• Master plan / preliminary design: Photorealistic or rendered plan view showing every plant and material
• Plant list / planting plan: Named species in a callout schedule
• Material schedule: Specific stones, woods, fixtures
• Section drawings: Vertical cuts showing grade changes, plantings, hardscape interfaces
• Cost estimate: Material takeoffs based on the master plan

Duration. 4–8 weeks.

AI workflow mapping. Design Development is the sweet spot for AI-assisted image editing — it's where image-based generative tools actually shine. The brief's three steps map cleanly:

| Brief Step | Maps to | |---------------------------|--------------------------------| | Add base plantings | Phase 4a (Base Plantings) | | (implied) Hardscape | Phase 4b (Hardscape Materials) | | Decor + finishing touches | Phase 4c (Decor) |

These are exactly the moves that make sense as layered edits on a rendered image, since the user is making aesthetic decisions on a visualized space, not yet specifying construction-grade dimensions.

Phase 5: Construction Documentation

What happens. The design is converted into a buildable contract document. Every dimension is specified. Every material is called out. Every detail is drawn. The contractor uses this set to bid and build.

Artifacts produced.

• Layout plans (dimensions, setouts)
• Grading and drainage plans
• Planting plan with quantities
• Lighting plan
• Irrigation plan
• Detail sheets
• Specifications book
• Permit application drawings

AI workflow mapping. Construction Documentation is completely out of scope for image-edit workflows. It requires precise dimensional data, code compliance, contractor-readable specifications, and detail drawings. The SimplyScapes Site Planner, Document Generation, and Hardscape Design capabilities address this layer separately. The AI tab does not need to reach into Phase 5.

---

2. Phases vs. AI-Assisted Image Editing — Mapping Table

| Phase | In Scope for AI Tab? | Mapped Brief Step | |-----------------------------|----------------------|-------------------------------------------------------------| | 1. Site Analysis | No (consumed only) | (Inputs from existing platform features; not a step here) | | 2. Programming | Yes — entry point | The lead-in questionnaire (goals, style, constraints) | | 3. Schematic Design | Yes | Step 1 (Prepare site) + Step 2 (Define spaces & layout) | | 4a. Base Plantings | Yes | Step 3 (Add base plantings) | | 4b. Hardscape Materials | Partial | Folded into Steps 2 and 3 depending on user goal | | 4c. Decor & Finishing | Yes | Step 4 (Decor + finishing touches) | | 4d. Detail Drawings | No | Out of scope — would require dimensional drawings | | 5. Construction Documents | No | Out of scope — handled by Site Planner and Document Gen |

---

3. The Step-Order Insight

Professional designers proceed top-down (largest, longest-lived, most-fixed first) through the design:

1. Bones first — trees, walls, terraces, major hardscape
2. Structure next — foundation shrubs, evergreen masses
3. Body — perennial beds, groundcovers, bulbs
4. Jewelry last — lighting, sculpture, accent pots, seasonal color

This is the inverse of how amateurs work. Amateurs put in the "jewelry" first (a fountain, a pot of annuals) because that's what caught their eye. The result is fragmented compositions that fight each other.

Implication for the AI tab. The proposed step order in the brief already follows the professional sequence:

1. Prepare the site (clear)
2. Define spaces and layout (bones)
3. Add base plantings (structure → body)
4. Decor + finishing (jewelry)
5. Views (camera angles, plant labels — not a design phase, just a review aid)

This sequencing is structurally correct by professional design standards. The challenge is making it feel intuitive to non-professionals who don't know they're working in this order.

---

4. The "Multiple Concepts" Pattern

A professional designer almost never presents a single design. They present 2–3 alternative concepts at the schematic phase, let the client choose, then refine the chosen concept. This pattern is so universal that interior-design tools (Decorilla, Modsy) have operationalized it: every customer is shown two or three concept boards before any further work happens.

Implication for the AI tab. The existing version-history / revert UI (called out as "preserved and elevated" in the brief) maps to this pattern. After the lead-in questionnaire, the AI tab can generate two or three concepts in parallel, present them as side-by-side thumbnails, and let the user pick a "main thread" to develop while keeping the others alive in the version tree.

---

5. The Style → System Insight

Professional designers don't pick "a style" and then design — they inherit a style vocabulary through training, regional norms, and client conversation, and then apply that vocabulary consistently across all decisions on a project.

The vocabulary is implicit but consistent:

• Plant palette: 5–10 core species, mass-planted, repeated
• Color palette: 3–5 hues that recur throughout
• Material palette: 2–3 primary hardscape materials
• Form language: Geometric (formal) or organic (informal)
• Density profile: Maximalist (cottage) or minimalist (modern)
• Maintenance posture: High-touch (formal) or low-touch (naturalistic)

A novice user has none of this implicit vocabulary. A pre-baked "design system" — a named bundle of these attributes the user can pick by name — gives the novice the same starting point a trained designer has. This is the central insight that makes the design- system-first approach work.

---

6. What This Means for the Brief

1. The five-phase model is the structural backbone of the AI tab workflow. The brief's existing five-step flow (prepare, layout, plant, decorate, views) is professionally correct.

2. The lead-in questionnaire is a Programming-phase artifact. It asks what's typically asked in the first client meeting: goals, uses, preferences, constraints.

3. The "design system" picker is a substitute for trained vocabulary. A novice picks "Cottage" and gets the same coherent vocabulary a trained designer would supply.

4. Multiple concepts are professional standard. The version-history UI is not a power-user feature — it's the schematic-phase concept review made tangible.

5. Phase 5 (Construction Documents) stays out of the AI tab. It belongs in Site Planner / Hardscape Design / Document Generation.

---

Sources

| # | Type | Reference | URL | |---|------|-----------|-----| | 1 | Standards | ASLA Documentation Standards | https://www.asla.org/2016awards/171889.html | | 2 | Extension | NC State Extension Gardener Handbook — Landscape Design | https://content.ces.ncsu.edu/extension-gardener-handbook/19-landscape-design | | 3 | Extension | UF/IFAS Basic Principles of Landscape Design (CIR536/MG086) | https://ask.ifas.ufl.edu/publication/MG086 | | 4 | Extension | UF/IFAS Landscape Design — Drawing a Planting Plan (EP456) | https://ask.ifas.ufl.edu/publication/EP456 | | 5 | Industry | Loft Six Four — Walk Through the Landscape Architecture Process | https://loftsixfour.com/blog/a-walk-through-the-landscape-architecture-process-what-developers-can-expect/ | | 6 | Industry | Verdance Landscape Architecture — Preliminary Design Phase | https://www.verdancedesign.com/blog/landscape-architecture-process-preliminary-design | | 7 | Industry | World Landscape Architect — Stages of a Landscape Architecture Project | https://worldlandscapearchitect.com/practice-stages-of-a-landscape-architecture-design-project/ | | 8 | Industry | The Landscape Library — Plant Symbols for Landscape Design | https://www.thelandscapelibrary.academy/blog/plant-symbols-for-landscape-design | | 9 | Industry | Bullard Bollards — 7 Landscape Design Principles | https://bullardbollards.com/7-landscape-design-principles-complete-guide-for-beautiful-outdoor-spaces/ |

---

Version History

| Date & Time (MT) | Author | Summary | |-------------------------|-------------------|----------------------------------------------------------------------------------------------------------| | 2026-05-01 03:30 PM MT | Research Analyst | Initial document. Five-phase ASLA model, mapping to AI tab steps, the bones-to-jewelry sequencing rule. |

AI Design Tools — Onboarding & Guided Workflow Patterns

Date: 2026-05-01 | Purpose: Document the onboarding flows of adjacent AI design tools — especially Anthropic's Claude Design — and extract patterns transferable to the Visual Designer AI tab. Audience: UX Designer, Product Lead.

TL;DR

Anthropic's Claude Design is the closest analogue to what SimplyScapes wants. It is the only major AI design tool that puts a design-system-first onboarding ahead of any creative work — users upload brand assets up front, Claude extracts a reusable system (colors, typography, components, layout), and every subsequent project inherits it automatically. The two-pane "chat + canvas" UI is now the dominant pattern for AI-assisted creation across Figma Make, Spacely AI, and Webflow AI. Interior-design services (Modsy, Havenly, Decorilla) and landscape services (Yardzen) all gate the experience behind a 5-minute style quiz that produces a "style profile" used as the seed for downstream design work. Vertical landscape competitors (Neighborbrite, Garden AI) skip the quiz and offer single-shot style transforms — the gap between their flow and a guided multi-step flow is exactly the differentiator SimplyScapes can own.

---

1. Anthropic Claude Design — Deep Dive

Claude Design is the closest analogue to the SimplyScapes vision — because it's the only major tool that puts a design-system-first flow ahead of any project work, then carries that system into every subsequent generation.

1.1 The Three-Stage Flow

Claude Design has three stages, in order:

1. Organization Onboarding (one-time) — set up the design system
2. Project Creation (each new project) — pick a fidelity, paste a prompt, optionally attach references
3. Iteration (continuous) — chat, comments, direct edits, sliders

The first stage is what differentiates it. Without it, the tool is a slightly nicer prompt-to-prototype. With it, every output is on-brand by default.

Sources: Set up your design system in Claude Design, Get started with Claude Design, Anthropic — Introducing Claude Design, DataCamp — What Is Claude Design?, Build Fast with AI — Claude Design Guide 2026, Claudia + AI — Starter Guide.

1.2 The Onboarding Flow (Design System Setup)

When a team first opens Claude Design, the onboarding takes them through these steps:

1. Create or switch organization — establishes a workspace where the design system lives.
2. Upload assets that define the brand — accepted inputs:
   • GitHub repository links (codebase context: components, tokens)
   • Local code files (uploaded zip)
   • Figma files
   • Font folders
   • Logo folders
   • Style notes (free-form text)
   • Slide decks or PDFs ("even a well-designed PowerPoint")
   • Screenshots of existing product or marketing
3. Claude reads and extracts — the system parses uploads and automatically generates a four-part design system:
   • Color palette: primary, secondary, accent
   • Typography: font families, sizes, weights
   • Components: buttons, cards, navigation, modals
   • Layout patterns: spacing, grid systems, page structures
4. Review the generated system — user inspects what Claude inferred and can edit before publishing.
5. Publish — flip the "Published" toggle. From this point, every new project in this organization automatically inherits the system.

Key UX observation. Claude Design does not ask the user a sequence of explicit questions ("What's your primary color? Pick from this list"). Instead, it asks for artifacts (files, links, references) and infers the answers. This is a deliberate choice — artifacts are what designers already have, and inference is more reliable than self-report. Users who don't have artifacts can paste text descriptions, but the tool nudges toward upload.

Why it matters for SimplyScapes. Most homeowners don't have a "brand artifact" — they have a yard photo. The SimplyScapes equivalent of "upload your brand assets" is a hybrid: the yard photo (already uploaded) plus a lightweight click-driven questionnaire plus a design-system pick (instead of system extraction). The user picks "Cottage" or "Modern" the way a Claude Design user picks an existing design system from their organization library.

1.3 The Project Creation Flow

Once the design system is published, each new project follows this shorter flow:

1. Click "New project" — pick from creation options shown in the left sidebar: Prototype, Slide deck, From Template, Other.
2. Pick a fidelity — Claude Design offers fidelity levels (e.g., "high fidelity" is documented; the full set is not publicly detailed). Higher fidelity costs more compute but produces more refined output.
3. Provide a prompt — text describing what you want. Anthropic's guidance instructs users to cover four elements:
   • Goal — what you're building
   • Layout — how things should be arranged
   • Content — what information to display
   • Audience — who will use it
4. Optionally attach references — upload images, sketches, web captures, documents. These act as visual seeds the model uses alongside the prompt.
5. Claude may ask clarifying questions — not always. When the prompt is too thin, Claude pauses and asks before rendering. When it has enough to proceed, it generates. Users who want to skip the clarifying step provide more upfront detail.
6. Canvas renders — the right-hand canvas shows the generated output. The left-hand chat panel stays open for refinement.

Key UX observation. The clarifying-questions behavior is conditional, not mandatory. This balances "guide me" novices against "let me cook" power users. It also avoids the trap of every adjacent tool: a 20-question intake survey before any output appears.

1.4 The Iteration Flow

After the canvas renders, four refinement channels are exposed simultaneously:

1. Chat panel (left) — for broad changes ("make it darker," "add a hero section," "regenerate as a landing page")
2. Inline comments on the canvas — user enters comment mode, clicks an element, types feedback ("Increase the horizontal padding on this button")
3. Direct text editing — fix copy without prompting
4. Tweak panel of sliders — Claude generates contextual sliders for spacing, color, layout — adjustable in real time

A documented quirk: comments occasionally disappear before Claude processes them. Workaround: paste the comment text into chat.

Key UX observation. Multiple refinement channels in parallel is the central insight. Different changes deserve different interaction patterns:

• Conceptual changes → chat
• Local element changes → inline comments
• Trivial copy fixes → direct edit
• Numeric adjustments → sliders

Forcing all changes through a single channel (text prompts) loses this granularity. The SimplyScapes AI tab can apply the same logic:

• "Reduce the lawn to a central oval" → chat or click affordance
• "Change this specific shrub to something taller" → inline action
• "Plant labels look wrong" → direct overlay toggle
• "Make the bed deeper" → slider

1.5 Export & Handoff

Claude Design's export menu offers PDF, HTML, PPTX, Canva handoff, and a special "Hand off to Claude Code" option that bundles the design intent for engineering pickup.

Key UX observation for SimplyScapes. The "hand off to Claude Code" pattern maps to "promote this AI design to a real Visual Designer project / save as proposal / send to MILO for refinement." The AI tab is a fast prototyping surface; the user should be able to graduate the result into the rest of the platform.

1.6 Outcomes and Speed

Reported outcomes from announcement and early customer interviews:

• Brilliant: complex pages went from "20+ prompts to recreate" in other tools to 2 prompts in Claude Design
• Datadog: teams went "from a rough idea to a working prototype before anyone leaves the room"
• General: work that used to take "a week of back-and-forth" now happens in single conversations

The speed advantage is largely the design-system inheritance — every output is on-brand without manual brand work. The same multiplier should apply to landscape design: a "Cottage" pick delivers a coherent vocabulary across plants, materials, color, and density without the user specifying each.

---

2. Adjacent AI Design Tools — Comparative Patterns

2.1 Figma Make — Two-Pane Pattern, Style Inheritance

Onboarding. Figma Make does not require a separate design-system setup — it inherits from your existing Figma library. You "add styling context from your Figma library to stay visually consistent with your design system."

Project creation. Four input modes:

1. Type a text prompt
2. Upload reference images or sketches
3. Upload documents (DOCX, PPTX, XLSX)
4. Web capture from live websites

Iteration. Same two-pane (chat + canvas) pattern Claude Design uses. Users select part of the design and prompt to refine.

Pattern transferable. Multi-modal input — text prompt, image upload, web capture. SimplyScapes already accepts photos; adding "reference image" upload (a Pinterest pin of an inspiration garden, for example) as a step-2 affordance is a natural extension.

Source: Figma Make Solutions — AI Onboarding App Builder, Figma Make landing page.

2.2 Spacely AI — Preset Mode vs. Prompt Mode

Spacely AI explicitly exposes two modes:

1. Preset Mode — guided experience. User selects from dropdowns for room type, style, lighting, materials. Click-driven.
2. Prompt Mode — describe vision in natural language.

Available styles: Modern, Minimalist, Japandi, Contemporary, Zen, plus spaces (Cafes, Restaurants, Hospitals, Offices, Buildings).

Iteration: adjust lighting, materials, furniture, layouts on top of existing work without rebuilding. Style transfer from a reference image to keep layout but apply a new aesthetic.

Pattern transferable. The mode toggle is structurally what SimplyScapes wants. The Visual Designer AI tab default is the SimplyScapes equivalent of Preset Mode (click-driven, guided steps). A paid-tier "type a prompt instead" feature is the Prompt Mode equivalent. Spacely AI proves the dual-mode UX is viable in production.

Source: Spacely AI — Tools, Spacely AI Review (10Web), Spacely AI on Medium — Revolutionizing Interior Design.

2.3 Modsy (defunct, but the operating model lives on at Decorilla and Havenly)

Onboarding. A 5-minute style quiz, mostly clicking on photos rather than answering text questions. The quiz output feeds an algorithm that produces a "style profile" used downstream.

Process.

1. Upload room photos from each corner
2. Take 5-minute style quiz (image-pick-driven)
3. AI builds a 3D model from photos using point-cloud reconstruction
4. Algorithm uses style profile + room model + product catalog to generate furnished room renderings
5. User can swap furniture in the 3D Editor (40 seconds per swap) using the Live Swap feature

Source: Modsy Review (Decorilla), Build an AI Interior Design App Like Modsy (IdeaUsher), Modsy on Digital Trends.

Pattern transferable. Photo-pick style quiz beats text quiz for taste capture. People can't tell you their style; they can recognize it. A 5-minute click-only quiz gathers more signal than a ten-question text intake. SimplyScapes can apply the same: show 8–12 image pairs ("which yard do you prefer?") and infer the style preference.

2.4 Havenly & Decorilla — Style Quiz + Designer Match

Both follow the same pattern as Modsy: style quiz → results page → designer match → upload photos → collaborate. The quiz is the single most important conversion artifact — it's what converts a visitor into someone who has a "design profile," which is what makes them likely to start a project.

The quiz returns a named style ("Bohemian," "Glam," "Midcentury Modern" — Havenly uses this terminology) that the user can identify with. This named style is the same construct as a SimplyScapes "design system" pick.

Source: Havenly Style Quiz, Havenly Blog — How to Find Your Design Style, Decorilla vs Havenly Comparison.

Pattern transferable. Naming the result matters. "Cottage" is a memorable shorthand the user identifies with for the rest of the session ("I want a Cottage with a vegetable garden"). Compare to a generic style profile vector — invisible and not human-readable.

2.5 Yardzen — Landscape-Specific Onboarding

The closest analogue in landscape design to Modsy's interior flow.

Nine-step onboarding.

1. Upload photos and videos of the property (multiple angles, slow-pan videos)
2. Wishlist questionnaire (functional needs)
3. Exterior inspiration quiz (style)
4. Goals definition (accessibility, slopes, special features)
5. Optional inspiration photo / Pinterest upload
6. Optional plans / drawings upload
7. Landscaping design style quiz
8. Plant preferences quiz
9. Submit

Total time: ~3 hours for a thorough customer; many take longer.

Source: Yardzen — How It Works, Yardzen Design Quiz, Yardzen Review (Lela Burris), Yardzen Style Guide (already cited in landscape-design-systems.md).

Pattern transferable.

• ✅ Multi-step intake works for landscaping. Users complete a 9-step process willingly.
• ✅ Style + plant preference are separate quizzes, not bundled.
• ❌ 3 hours is too long for the AI tab (Yardzen is a paid concierge service; the AI tab needs to deliver value in <5 minutes).

The right calibration: take Yardzen's information categories (property photos, wishlist, exterior style, goals, inspiration, plant preferences) but compress them to 4–6 click steps by using image-driven multi-select and inferring more from the photo.

2.6 Neighborbrite & Garden AI — Vertical AI Landscape Competitors

Neighborbrite flow (free version).

1. Upload yard photo
2. Pick from 16 garden styles (Mediterranean, fairy garden, edible garden, etc.)
3. Generate (~30–60 seconds)

Neighborbrite paid: "Create your style"

• Refine designs by telling the AI what to add or remove
• Add unique elements via reference images
• Select an area (flowerbed, lawn, walkway) for partial re-render
• Plant lists tailored to location, hardiness zone, sunlight

Source: Neighborbrite landing, Neighborbrite for Landscapers, Neighborbrite Review (DeClom), Neighborbrite (AI Tools Reviewer).

Pattern transferable.

• ✅ "Pick a style → generate" is the dominant landscape AI pattern. Users expect it to be available.
• ✅ Partial re-render (select a region, change just that part) is a must-have.
• ❌ The flow is single-shot — no programming phase, no schematic phase, no iteration toward a more refined design.

This is the differentiation gap. Neighborbrite shows users expect "pick a style." SimplyScapes can satisfy that expectation in 30 seconds (the fast lane: pick a system, see the result) and then offer the multi-step refinement path that Neighborbrite doesn't have.

2.7 Higharc — Architectural Multi-Step (B2B Reference)

Higharc is for homebuilders, not homeowners — but the pattern is useful as an enterprise reference.

Flow. Upload sketch or floor plan → AI detects rooms → AI populates with doors, windows, fixtures → 3D model generated → buyer runs a sales configurator to customize options.

The interesting move is the handoff between AI-generated baseline and configurator-driven customization. The AI sets up the baseline; the human picks options on top of it.

Source: Higharc — Studio, Higharc — From Sketch to 3D Model in Minutes, Higharc — AI Capabilities Announcement.

Pattern transferable. AI-generated baseline + human-driven configurator on top is a strong pattern. SimplyScapes can do the same: pick a design system → AI generates a "starting render" with the system applied → user configures specific plants and elements on top using the existing object library + drag/drop.

2.8 Maket — AI Floor Plan Configurator

Flow. Specify room types + square footage + architectural shape + design style → AI generates multiple variations → user compares and customizes.

Source: Maket, Maket Features.

Pattern transferable. The "specify constraints, get multiple variations" pattern overlaps with the schematic-design phase (multiple concepts). SimplyScapes already plans to use version history for this — Maket validates that the multiple-variations pattern is now expected in AI design tools.

---

3. Cross-Tool Pattern Synthesis

| Pattern | Tools that use it | Apply to SimplyScapes? | |----------------------------------------------------|--------------------------------------------------------|------------------------| | Design-system-first onboarding | Claude Design | Yes — central | | Two-pane chat + canvas | Claude Design, Figma Make, Spacely AI | Yes (already shipped) | | Preset Mode vs. Prompt Mode toggle | Spacely AI | Yes — paid tier | | Image-pick style quiz (photos, not text) | Modsy, Havenly, Decorilla | Yes — lead-in | | Named style outputs ("Cottage", "Modern") | Havenly, Decorilla, Yardzen, Neighborbrite | Yes — design systems| | Multi-step intake (5+ screens) | Yardzen | Compress to 4–6 | | Single-shot style transform | Neighborbrite, Garden AI | Offer as fast lane | | Partial re-render (select region) | Neighborbrite, Spacely AI | Yes (already exists) | | Multiple refinement channels (chat, comments, sliders) | Claude Design, Figma Make | Yes — adopt | | AI baseline + human configurator on top | Higharc, Maket | Yes — drag/drop | | Multiple concept variations in parallel | Maket, Decorilla, Modsy, professional designers | Yes — version tree | | Clarifying questions only when prompt is thin | Claude Design | Yes — conditional |

---

4. The Onboarding Spectrum

AI design tools sit on a spectrum from "no onboarding" to "full design system setup":

None                                                                Full
|---------------------|---------------------|---------------------|
                      |                     |                     |
              Garden AI              Modsy/Havenly           Claude Design
              Neighborbrite          Yardzen                 (full org-level
              (free version)         (5–60 min quiz)          system, asset
              (no quiz)                                       extraction)
                      |                     |                     |
              "Pick a style,         "Style quiz drives     "Upload assets,
               see a result"          designer match"        we extract
                                                             system, every
                                                             output is
                                                             on-brand"

Where SimplyScapes should sit. Between Modsy/Havenly and Claude Design. More structured than Neighborbrite (because pros need a real workflow, not a one-shot transform), less heavy than Yardzen (because we need <5-minute time-to-first-render). The pre-baked design system picker is the SimplyScapes innovation that sits in this gap.

---

5. The "Default State Must Be Useful" Constraint

A central constraint from the brief: "Default state must be immediately useful with zero typing."

How adjacent tools solve this:

• Neighborbrite: Default state shows a grid of 16 named styles the user clicks. Zero typing. Immediate value.
• Spacely AI Preset Mode: Default state shows dropdown menus pre-filled with sensible defaults. User changes what they want, generates. Zero typing.
• Yardzen: Default state is the upload step. User uploads photos and proceeds through clicks.
• Modsy: Default state is the style quiz launcher. User clicks "Take quiz" and is in image-picking mode within seconds.
• Claude Design: Default state is the project creation modal with a "Prototype" / "Slide deck" / "From Template" picker. Some typing required — Claude Design is positioned for designers, not novices.

Pattern. Every tool except Claude Design solves "useful by default" with a click-driven picker that shows a list of named named outputs (styles, room types, templates). The user's first action is a click, not a keystroke. The SimplyScapes AI tab default state should be the same shape: a grid of named design systems with preview images, plus an obvious "use my last design" / "start from a photo" affordance.

---

6. What This Means for the AI Tab

1. Build the design-system picker as the lead-in, the way Claude Design builds the org-level design system. The selection happens once per session and propagates through every subsequent step.

2. Use a click-driven, image-pick lead-in questionnaire modeled on Modsy / Havenly / Yardzen. Show paired photos. Infer style from picks rather than asking text questions.

3. Adopt the Spacely-AI dual-mode pattern. Default = Preset Mode (click-driven). Pro/paid tier = Prompt Mode (free text). The toggle is visible but the default is clearly the right choice.

4. Borrow the Claude Design four-channel iteration model. Chat panel for broad changes, inline element controls for local edits, direct overlay toggles for trivial fixes, sliders for numeric tweaks. Don't force everything through chat.

5. Adopt the Higharc / Maket "AI baseline + configurator" hybrid. AI generates the design-system-flavored baseline render. User drags and drops elements from the existing object library on top. "Apply" flattens the staged elements into the next render.

6. Multiple concepts via version history is professional standard, not a power-user feature. Decorilla, Modsy, and every professional designer present 2–3 concepts. Surface the existing version history as a first-class concept-comparison surface.

7. Default state shows named systems — a grid of 8–12 design- system tiles with hero images. Zero typing required to start.

---

Sources

| # | Type | Reference | URL | |---|------|-----------|-----| | 1 | Help Doc | Claude Design — Set up your design system | https://support.claude.com/en/articles/14604397-set-up-your-design-system-in-claude-design | | 2 | Help Doc | Claude Design — Get started | https://support.claude.com/en/articles/14604416-get-started-with-claude-design | | 3 | Press | Anthropic — Introducing Claude Design | https://www.anthropic.com/news/claude-design-anthropic-labs | | 4 | Review | DataCamp — What Is Claude Design? | https://www.datacamp.com/blog/claude-design | | 5 | Review | Build Fast with AI — Claude Design Guide 2026 | https://www.buildfastwithai.com/blogs/claude-design-anthropic-guide-2026 | | 6 | Review | Claudia + AI — Claude Design Starter Guide | https://claudiaplusai.substack.com/p/claude-design-starter-guide-and-examples | | 7 | Review | MindStudio — What Is Claude Design? | https://www.mindstudio.ai/blog/what-is-claude-design-anthropic-design-tool | | 8 | Solutions | Figma Make — AI Onboarding App Builder | https://www.figma.com/solutions/ai-onboarding-app-builder/ | | 9 | Product | Figma Make | https://www.figma.com/make/ | | 10 | Product | Spacely AI | https://www.spacely.ai/ | | 11 | Review | Spacely AI Review (10Web) | https://10web.io/ai-tools/spacely-ai/ | | 12 | Blog | Spacely AI — Revolutionizing Interior Design | https://medium.com/spacelyai/revolutionizing-interior-design-with-ai-an-introduction-to-spacely-ai-a381a85a9efd | | 13 | Review | Modsy Review (Decorilla) | https://www.decorilla.com/online-decorating/modsy-review/ | | 14 | Article | How Modsy Works (Digital Trends) | https://www.digitaltrends.com/home/modsy-interior-design/ | | 15 | Tutorial | Build an AI Interior Design App Like Modsy | https://ideausher.com/blog/build-ai-interior-design-app-modsy/ | | 16 | Quiz | Havenly Interior Design Style Quiz | https://havenly.com/interior-design-style-quiz | | 17 | Blog | Havenly — How to Find Your Design Style | https://havenly.com/blog/find-your-design-style | | 18 | Comparison | Decorilla vs Havenly | https://www.decorilla.com/online-decorating/decorilla-vs-havenly-which-online-interior-design-service-is-right-for-you/ | | 19 | Process | Yardzen — How It Works | https://yardzen.com/how-it-works | | 20 | Quiz | Yardzen Design Quiz | https://yardzen.com/design-consultation/design-quiz | | 21 | Review | Yardzen Review (Lela Burris) | https://www.lelaburris.com/yardzen-review/ | | 22 | Product | Neighborbrite | https://neighborbrite.com/ | | 23 | Product | Neighborbrite for Landscapers | https://neighborbrite.com/landscapers | | 24 | Review | Neighborbrite Review (DeClom) | https://declom.com/neighborbrite/ | | 25 | Product | Higharc — Studio | https://www.higharc.com/product/studio | | 26 | Blog | Higharc — From Sketch to 3D Model in Minutes | https://www.higharc.com/blog/homebuilding-ai-3d-modeling-higharc | | 27 | Press | Higharc — AI Capabilities Announcement | https://www.higharc.com/newsroom/higharc-announces-new-ai-capabilities-for-industry-leading-homebuilding-platform | | 28 | Product | Maket — AI Floor Plan Creation | https://www.maket.ai/ | | 29 | Features | Maket — Features | https://www.maket.ai/features |

---

Version History

| Date & Time (MT) | Author | Summary | |-------------------------|-------------------|--------------------------------------------------------------------------------------------------------------------| | 2026-05-01 04:00 PM MT | Research Analyst | Initial document. Deep dive on Claude Design + comparative scan of Figma Make, Spacely AI, Modsy, Havenly, Decorilla, Yardzen, Neighborbrite, Higharc, Maket. |

Landscape Design Systems — Translating the Web Design System Concept

Date: 2026-05-01 | Purpose: Translate the "design system" concept (familiar from web/UI design) to landscape design, propose a coherent attribute model, and propose a list of named pre-baked systems that ship with the product. Audience: UX Designer, Product Lead.

TL;DR

A landscape "design system" is a coherent vocabulary of attributes that, when chosen up front, produces consistent decisions across every subsequent step of the design — plant choices, hardscape materials, form language, density, color, and maintenance posture. The eight attributes worth controlling are plant palette, hardscape vocabulary, formality, density, color/seasonality, maintenance load, water use, and form language. We recommend shipping 12 pre-baked systems covering the major recognized styles (Cottage, Modern, Naturalistic, Xeriscape, Mediterranean, Japanese Zen, Tropical, Prairie/Native, Formal Traditional, Modern Farmhouse, Coastal, Rustic). Three are "essentials" the user picks first; the other nine are organized by climate fit.

---

1. The Web Design System Analogy

In web/UI design, a "design system" is a coherent set of decisions about:

• Color — primary, secondary, accent palettes, plus dark/light variants
• Typography — font families, sizes, weights, line heights
• Spacing — base unit, multipliers, padding/margin scale
• Components — buttons, inputs, cards, navigation, modals
• Patterns — page layouts, form layouts, dialog patterns
• Motion — easing curves, durations, choreography rules

The system propagates downward: once the design system is chosen, every component decision becomes a constrained pick from the system rather than a free choice. The benefit is consistency without constant judgment-calls.

Anthropic's Claude Design uses this exact pattern: the user uploads brand assets, Claude extracts a system, and every project after that inherits the system. Every output is on-brand by default.

---

2. The Landscape Equivalent — Eight Attribute Categories

Translating these categories to landscape design produces eight attribute groups. Each one maps to a specific generative-AI controllable parameter or prompt component.

2.1 Plant Palette

The set of plant species (or species-group archetypes) that populate the design.

Sub-attributes.

• Core species (5–10): the plants that recur throughout. Mass- planted in groups of 3–7, repeated for cohesion.
• Accent species (3–5): focal plants — specimen trees, structural evergreens, dramatic perennials.
• Filler/groundcover species (2–4): the matrix plants that connect beds.
• Plant family bias: e.g., "ornamental grasses dominant," "broadleaf evergreens dominant," "deciduous flowering shrubs dominant."

Sources.

• University of Florida MG086 — Basic Principles of Landscape Design: "Most properties [use] 20–30 core plants used repeatedly throughout the landscape with perhaps another 10–15 special accent plants for focal areas."
• The Landscape Library — Plant Symbols: professional firms maintain CAD libraries of plant blocks structured exactly this way.

2.2 Hardscape Vocabulary

The materials and forms of the non-living elements — paving, walls, edging, structures.

Sub-attributes.

• Primary material (1–2): the dominant pavers/stone (e.g., bluestone, decomposed granite, pea gravel, brick).
• Secondary material (1–2): wood, metal, concrete accents.
• Edge treatment: clean cut, mortared, dry-laid, mounded soil, metal edge.
• Vertical structures: pergola, arbor, fence (wood vs. metal vs. stone).
• Water/fire features: fountain, pond, fire pit, fire bowl, none.

Pattern source: "A unified material palette for each project typically includes 2–3 primary hardscape materials that repeat throughout the landscape".

2.3 Formality

A scale from highly geometric/symmetrical to organic/wild.

Scale.

1. Formal — symmetrical, geometric, clipped, axial
2. Semi-formal — geometric layout with looser planting
3. Transitional — mix of geometric and organic forms
4. Informal — curved beds, asymmetric massing
5. Naturalistic — wild, plant-driven, no apparent geometry

This scale maps directly to UF/IFAS's line and form discussion: "Geometric forms — circles, squares, and clean polygons — produce a formal, structured look... Naturalistic forms include meandering, organic, fragmented edges."

2.4 Density

How visually dense the plantings are — sparse to lush.

Scale.

1. Minimal — wide negative space, single specimens, gravel matrices (Modern, Xeriscape, Japanese)
2. Moderate — standard residential density (Modern Farmhouse, Coastal, Mediterranean)
3. Lush — full beds, layered plantings (Cottage, Tropical, English)

2.5 Color & Seasonality

The color palette across seasons.

Sub-attributes.

• Color palette: the 3–5 hues that recur. Examples:
  • Whites / silvers / pale blues (Formal Traditional, Moditerranean)
  • Bright multi-colored (Cottage)
  • Earth tones / rust / amber (Prairie, Mediterranean)
  • Deep greens with bursts of bold color (Tropical)
  • Cool grays + greens + neutrals (Modern Minimal)
• Seasonality: spring-dominant, multi-season, evergreen-dominant, fall/winter interest emphasized.

2.6 Maintenance Load

The hours-per-month a homeowner or pro can expect to spend.

Scale.

1. Very low (≤2 hrs/month): Xeriscape, Modern Minimal, Native/Prairie post-establishment
2. Low (3–5 hrs/month): Naturalistic, Coastal, Rustic, Modern Farmhouse
3. Moderate (6–10 hrs/month): Mediterranean, Modern, Japanese, French Country
4. High (10–20 hrs/month): Cottage, Formal Traditional, Tropical
5. Very high (>20 hrs/month): Estate-formal, English garden with roses, large lawn programs

2.7 Water Use

Annual irrigation requirement.

Scale.

1. Xeric (no irrigation after establishment): Xeriscape, Desert, Native Prairie
2. Low (drip + occasional): Mediterranean, Naturalistic, Modern Farmhouse, Modern Minimal, Coastal
3. Moderate (regular drip): Modern, Japanese, Formal Traditional
4. High (regular spray + drip): Cottage, English, Tropical

2.8 Form Language

The vocabulary of shapes that recur in plant masses, beds, and hardscape.

Categories.

• Geometric — straight lines, circles, squares, clipped hedges (Formal Traditional, Modern Minimal)
• Curvilinear — flowing curves, soft edges (Cottage, English, Coastal)
• Architectural — strong vertical/horizontal planes, clean rectangles (Modern, Modern Farmhouse)
• Organic — irregular, naturalistic, terrain-following (Naturalistic, Prairie, Native)
• Cultural-specific — referenced from a specific tradition (Japanese rocks-and-water, French parterres)

---

3. The Attribute Matrix — How Attributes Combine

Each design system specifies a value across all eight attributes. The matrix below shows how the attributes vary across the proposed 12 systems. A "design system" is essentially a row in this matrix — a coherent set of values across all eight axes.

| System | Plant Palette Tone | Hardscape | Formality | Density | Color | Maintenance | Water | Form Language | |-------------------|------------------------|---------------------|-------------|-----------|------------------------|-------------|----------|------------------| | Cottage | Lush perennials, mixed | Brick/curved stone | Informal | Lush | Multi-colored bright | High | Moderate | Curvilinear | | Modern | Architectural plants | Concrete, large pavers | Semi-formal | Moderate | Greens, grays, whites | Moderate | Low | Architectural | | Naturalistic | Native masses | Minimal hardscape | Naturalistic| Moderate | Earth tones, fall hues | Low | Low | Organic | | Xeriscape | Drought-tolerant | Decomposed granite | Transitional| Minimal | Tans, silvers, sage | Very low | Xeric | Architectural | | Mediterranean | Lavender, olive, sage | Stone, gravel | Semi-formal | Moderate | Silvers, terracotta | Moderate | Low | Curvilinear | | Japanese Zen | Maples, moss, pines | Stone, gravel, wood | Semi-formal | Minimal | Greens, earth tones | Moderate | Moderate | Cultural-specific| | Tropical | Large-leafed bold | Wood deck, stone | Informal | Lush | Deep greens, bold | High | High | Organic | | Prairie / Native | Grasses, wildflowers | None or stone path | Naturalistic| Moderate | Earth tones, gold | Low | Xeric | Organic | | Formal Traditional| Boxwood, hydrangea | Bluestone, brick | Formal | Moderate | Whites, greens | High | Moderate | Geometric | | Modern Farmhouse | Grasses, perennials | Decomposed granite, herringbone brick | Transitional | Moderate | Light greens, silvers | Low | Low | Architectural | | Coastal | Grasses, billowy | Gravel, rustic wood | Informal | Moderate | Driftwood, blues | Low | Low | Curvilinear | | Rustic | Full textural plants | Wood, mulch, gravel | Informal | Moderate | Warm wood tones | Moderate | Low | Organic |

Source for style data: Yardzen — A Guide to Landscaping Styles, Lamacchia Landscape Construction — Popular Landscape Design Styles, Cedreo — Common Landscape Design Styles, Pretty Purple Door — Garden Styles, NC State Extension — Landscape Design.

---

4. The Twelve Recommended Pre-Baked Systems

We propose shipping 12 named systems organized into three groups of four. The grouping reflects how users self-identify — by life- stage and lifestyle, not by climate. Climate filtering happens behind the scenes based on the property's hardiness zone.

Group A: The Essentials (universally recognized, high pick rate)

These four are the styles a typical homeowner would name unprompted when asked "what kind of yard do you want?" Yardzen's data and Houzz's catalog both put these at the top.

1. Cottage

Lush, romantic, multi-colored. Roses climbing on pickets, lavender spilling onto curved brick paths, perennials packed shoulder-to- shoulder. Maintenance is real but the result is the most emotionally rewarding garden style in the catalog.

Starter attributes. Plant palette: roses, lavender, salvia, foxglove, hollyhock, hydrangea, climbing vines. Hardscape: curved brick paths, weathered wood, picket fence. Formality: Informal. Density: Lush. Color: bright multi-colored. Maintenance: High (10–20 hrs/mo). Water: Moderate. Form: Curvilinear.

2. Modern

Clean, calm, intentional. Negative space carries as much weight as the plantings. A few architectural plants (agave, ornamental grass drifts, columnar evergreens) anchor wide expanses of paving or gravel. Reads as deliberate luxury.

Starter attributes. Plant palette: agave, ornamental grasses (Calamagrostis, Miscanthus), columnar evergreens, white salvia, boxwood orbs. Hardscape: concrete, large-format pavers, steel edging. Formality: Semi-formal. Density: Moderate. Color: Greens, grays, whites. Maintenance: Moderate. Water: Low. Form: Architectural.

3. Naturalistic

Looks wild but is engineered. Native and adapted plants in masses that mimic local ecosystems — meadow drifts, woodland edges, shrubland. Pollinator-rich. The lowest-maintenance "real garden" style once established.

Starter attributes. Plant palette: native grasses (Bouteloua, Schizachyrium, Panicum), echinacea, monarda, rudbeckia, native shrubs. Hardscape: minimal — gravel paths, stone slab steppers. Formality: Naturalistic. Density: Moderate. Color: Earth tones, fall colors. Maintenance: Low. Water: Low. Form: Organic.

4. Formal Traditional

Geometric, symmetrical, classic. Clipped hedges, axial paths, symmetrical perennial beds, white-and-green color discipline. The "estate garden" look — works best on regular lots with a strong architectural home.

Starter attributes. Plant palette: boxwood, yew, hydrangea, hosta, white roses, white peonies. Hardscape: bluestone pavers, brick borders, formal urns. Formality: Formal. Density: Moderate. Color: Whites, greens, pale blues. Maintenance: High. Water: Moderate. Form: Geometric.

Group B: Regional / Climate-Driven (climate-fit recommendations)

These four are climate-anchored and become defaults in the matching zones, but available everywhere.

5. Xeriscape

Drought-proof. Climate-adapted plants over decomposed granite or permeable gravel, with very limited turf or none. Minimal irrigation after establishment. The right answer for most of the American West and increasingly for water-restricted regions nationally.

Starter attributes. Plant palette: agave, yucca, lavender, sage, penstemon, ornamental grasses, native succulents. Hardscape: decomposed granite, river rock, boulder accents. Formality: Transitional. Density: Minimal. Color: Tans, silvers, sages. Maintenance: Very low. Water: Xeric. Form: Architectural.

6. Mediterranean

Warm-region classic. Lavender, olive, rosemary, citrus. Stone walls, gravel courtyards, terracotta accents, pergolas with climbers. Made for heat and dry summers but adaptable to many climates.

Starter attributes. Plant palette: lavender, rosemary, olive (where hardy), citrus, salvias, rockrose. Hardscape: limestone or travertine pavers, gravel, terracotta pots, pergolas. Formality: Semi-formal. Density: Moderate. Color: Silvers, terracotta, deep greens. Maintenance: Moderate. Water: Low. Form: Curvilinear.

7. Tropical

Lush, bold, escapist. Large-leafed plants, vivid color, dense layering. Built for warm-climate properties (Zones 9–11) but a "containerized tropical" version is achievable in colder zones.

Starter attributes. Plant palette: bird of paradise, hibiscus, bougainvillea, palms, banana, philodendron, cordyline. Hardscape: wood decking, stone, water features. Formality: Informal. Density: Lush. Color: Deep greens with vivid color bursts. Maintenance: High. Water: High. Form: Organic.

8. Coastal

Salt-tolerant, breezy, weathered. Ornamental grasses, lavender, rosemary, billowy plantings against weathered wood and gravel. The look people want even when not actually on the coast.

Starter attributes. Plant palette: beach grasses, lavender, sea lavender, hydrangea, agapanthus, salt-tolerant shrubs. Hardscape: gravel, weathered wood, rope details. Formality: Informal. Density: Moderate. Color: Driftwood, blues, sandy neutrals. Maintenance: Low. Water: Low. Form: Curvilinear.

Group C: Specialty / Cultural / Lifestyle

These four are recognizable, named, and meaningful — but pickier choices that match specific cultural or lifestyle goals.

9. Japanese Zen

Contemplative, restrained, deliberate. Mossy ground, raked gravel, placed stones, Japanese maples, pines. Every element is intentional. The opposite of cottage maximalism.

Starter attributes. Plant palette: Japanese maple, black pine, moss, ferns, hostas, bamboo (clumping). Hardscape: raked gravel beds, placed stones, stepping stones, lanterns, bamboo screens. Formality: Semi-formal. Density: Minimal. Color: Greens, earth tones, accents of red. Maintenance: Moderate. Water: Moderate. Form: Cultural-specific.

10. Modern Farmhouse

Modern's quieter cousin. Grasses, soft perennials, decomposed granite paths, split-rail fencing, herringbone brick. Pairs with board-and-batten or shingled architecture. Comfort meets discipline.

Starter attributes. Plant palette: yarrow, Russian sage, artemisia, ornamental grasses, redbud, hydrangea. Hardscape: herringbone brick, decomposed granite, split-rail fence, wood furniture. Formality: Transitional. Density: Moderate. Color: Light greens, silvers, pale blues. Maintenance: Low. Water: Low. Form: Architectural.

11. Prairie / Native

Habitat-first. Tall grasses and native wildflowers in great drifts, moving with the wind. Almost no hardscape. The most ecologically impactful style — a designed pollinator preserve.

Starter attributes. Plant palette: big bluestem, little bluestem, switchgrass, echinacea, monarda, asters, goldenrod, native shrubs. Hardscape: minimal — mown grass paths or stone slab steppers. Formality: Naturalistic. Density: Moderate. Color: Earth tones, gold, rust. Maintenance: Low. Water: Xeric. Form: Organic.

12. Rustic

Warm, textural, organic. Wood, mulch, gravel, full plantings with emphasis on texture and form rather than bloom color. Pairs with log cabin, bungalow, and craftsman architecture.

Starter attributes. Plant palette: textural shrubs, ferns, hostas, ornamental grasses, native trees. Hardscape: wood (rough-sawn or split rail), mulch, gravel, fieldstone. Formality: Informal. Density: Moderate. Color: Warm wood tones, deep greens. Maintenance: Moderate. Water: Low. Form: Organic.

---

5. Mapping a Design System Pick to a Generative Prompt

Each design system is, in production, a structured prompt fragment. When the user picks "Cottage," the AI tab assembles this into the prompt for downstream image generation:

Style: English Cottage Garden
Plant palette: roses, lavender, salvia, foxglove, hydrangea, climbing
vines; emphasis on multi-colored perennials in dense plantings.
Hardscape: curved brick paths, weathered wood structures,
picket-style fencing.
Color palette: bright multi-colored with whites and pinks dominant;
warm wood tones for hardscape.
Formality: informal — curving bed shapes, organic massing,
no straight lines.
Density: lush — beds packed with overlapping plantings.
Maintenance: high — mature, established garden look.
Form language: curvilinear — flowing curves, soft edges.

This fragment is concatenated with the per-step intent ("add base plantings," "remove that overgrown shrub") to produce the final generation prompt. The design system is the constant context; the step intent is the variable.

---

6. The User's Mental Model

The user does not need to understand the eight-attribute taxonomy. They pick one named system ("Cottage") and the eight attributes are pre-set. This is exactly how Claude Design works — the user picks a design system, all the underlying tokens (colors, typography, spacing) come along for free.

Where the system breaks down:

• User wants two systems blended ("Cottage but lower maintenance"). This becomes a UX challenge — do we offer "blend two systems" or do we add modifiers ("low-maintenance variant")? Recommend the modifier path for v1; blending is a later feature.
• User's home doesn't match the system they picked (e.g., Tropical on a 1920s craftsman in Minnesota). The questionnaire step should soft-flag mismatches without blocking — "Tropical works best in Zones 9–11; your zone is 4. Want to try Modern Farmhouse instead?"

---

7. Why Twelve, Not Twenty

Yardzen ships 21 styles. We propose 12 deliberately:

• Cognitive load. A 4×3 grid is easier to scan than a 5×5 grid.
• Decision fatigue. Past 8–10 choices, users feel paralyzed rather than empowered.
• Authoring cost. Each system requires a polished hero image, a starter prompt, and tested attribute values. Twelve well-tuned systems produce better outputs than twenty rough ones.
• The 80/20 rule. Yardzen's 21 styles include "Family-Friendly," "Edible Garden," "Transitional," "Standard Traditional" — these are mostly modifiers or hybrids of the core 12.
• Room to grow. Holding back 8–10 systems for v2 (English, French Country, Southwestern, Desert, Modern Bohemian, Modern Minimal, Moditerranean, Craftsman, Edible Garden) keeps a roadmap.

---

8. The "Modifier" Concept (Future)

The eight-attribute taxonomy lets us layer modifiers on top of base systems without combinatorial explosion:

• "Lower maintenance" — slides Maintenance attribute one rung toward Low across whatever base system is selected
• "Family-friendly" — adds turf, shifts plant choices to non-thorny / non-toxic
• "Pollinator-focused" — biases plant palette toward natives and known pollinator plants
• "Edible" — overlays an edible-plant track on the base palette (raised veggie beds, fruit trees, herb spirals)

Modifiers are deferred to v2 but the attribute model supports them.

---

9. What This Means for the Brief

1. Twelve systems is the recommended catalog. Three groups of four organize the picker neatly into a 4×3 grid.
2. Each system maps to a structured prompt fragment with values across all eight attributes. The fragment is the prompt context constant for the rest of the session.
3. The eight-attribute taxonomy is the internal data model. The user never sees it explicitly; it is the schema underneath.
4. A pre-built example image per system is non-negotiable. Names alone don't communicate "Cottage" the way a photo does.
5. Climate compatibility is a soft-flag, not a hard block. The system can offer "consider X instead" without preventing the pick.
6. The starter attribute values in this doc are seeds for the spec phase. Real production values will need horticultural review (climate fit, regional plant substitutions, real plant DB lookups) before launch.

---

Sources

| # | Type | Reference | URL | |---|------|-----------|-----| | 1 | Extension | UF/IFAS Basic Principles of Landscape Design (CIR536/MG086) | https://ask.ifas.ufl.edu/publication/MG086 | | 2 | Extension | NC State Extension Gardener Handbook — Landscape Design | https://content.ces.ncsu.edu/extension-gardener-handbook/19-landscape-design | | 3 | Industry | Yardzen — A Guide to Landscaping Styles | https://yardzen.com/yzblog/landscaping-ideas-for-every-style | | 4 | Industry | Lamacchia Landscape Construction — Popular Landscape Design Styles | https://lamacchialandscapeco.com/landscape-design-styles/ | | 5 | Industry | Cedreo — Common Landscape Design Styles | https://cedreo.com/blog/landscape-design-style-guide/ | | 6 | Industry | Pretty Purple Door — Garden Styles | https://www.prettypurpledoor.com/garden-styles/ | | 7 | Industry | Bullard Bollards — 7 Landscape Design Principles | https://bullardbollards.com/7-landscape-design-principles-complete-guide-for-beautiful-outdoor-spaces/ | | 8 | Reference | The Landscape Library — Plant Symbols for Landscape Design | https://www.thelandscapelibrary.academy/blog/plant-symbols-for-landscape-design | | 9 | Magazine | HGTV — What Is an English Garden | https://www.hgtv.com/outdoors/gardens/garden-styles-and-types/english-garden-design | | 10 | Magazine | Houzz — Lay of the Landscape: English-Style Gardens | https://www.houzz.com/magazine/lay-of-the-landscape-english-style-gardens-stsetivw-vs~1822380 | | 11 | Reference | Landscaping Network — A Complete Collection of Landscape Styles (PDF) | https://www.landscapingnetwork.com/garden-styles/design-guide.pdf | | 12 | Reference | Landscaping Network — Xeriscape Fundamentals | https://www.landscapingnetwork.com/landscape-design/Xeriscape.html | | 13 | Industry | Outside Landscape Group — Different Types of Landscaping | https://outsidelandscapegroup.com/different-types-of-landscaping/ | | 14 | Reference | Penn State Extension — Principles of Garden Design | https://extension.psu.edu/principles-of-garden-design | | 15 | Magazine | Garden Design — Tropical Garden Ideas | https://www.gardendesign.com/tropical/ideas.html | | 16 | Magazine | Monrovia — Garden Color Palette Guides | https://www.monrovia.com/be-inspired/garden-color-palette-guides.html | | 17 | Extension | Illinois Extension — Naturalistic Landscapes | https://extension.illinois.edu/blogs/good-growing/2025-08-15-nature-design-naturalistic-landscapes |

---

Version History

| Date & Time (MT) | Author | Summary | |-------------------------|-------------------|------------------------------------------------------------------------------------------------------------------------| | 2026-05-01 04:30 PM MT | Research Analyst | Initial document. Eight-attribute landscape design system model + 12 named pre-baked systems with starter attribute values. |

Home Architectural Style and Existing Landscape Style — Classification State of the Art

Date: 2026-05-01 | Status: draft Parent report: research/report.md Part VI Idea: generative-ai-chat-interface

This document covers Deltas 1 and 2 of the 2026-05-01 research round — both reduce to the same core technique (vision-based aesthetic classification of an outdoor scene) and so are treated together. The goal is a workable, ship-soon classifier the AI panel can use to:

1. Auto-detect the home's architectural style at site-analysis time (Craftsman, Ranch, Modern, etc.) — used to soft-recommend matching design systems on the picker and to seed questionnaire defaults.
2. Auto-detect the existing landscape's style in the yard photo (Cottage-leaning, Bare lawn, Naturalistic, Overgrown, etc.) — used to inform the "what to keep" decision in Step 1 (Prepare site).

Both auto-classifications are confirmable / correctable. They are starting points, not verdicts.

---

A. Home Architectural Style Classification

A.1 Available pretrained models and open classifiers

The academic baseline is the Architectural Style dataset (also called the Xu/ECCV 2014 dataset), which contains roughly 4,799 images across 25 architectural styles, including American Craftsman, American Foursquare, Colonial, Tudor Revival, Mid-Century Modern, and many world-architecture categories. The dataset is widely cited but publicly available image counts are uneven across classes (some classes have ~50 images). The original Xu et al. paper, "Architectural Style Classification Using Multinomial Latent Logistic Regression", reported 46% top-1 accuracy on the 25-class problem and 69% on a 10-class subset.

Modern CNN approaches push numbers higher. The 2022 channel-spatial attention CNN paper on the same dataset reports significant gains over the 2014 baseline, though still in the 70-80% range on the 25-class problem. The Roboflow ArchiStyles dataset provides a modern pretrained classifier with 25 categories, ~300 sample images per category, and a 70/20/10 train/val/test split.

The most directly relevant prior art for SimplyScapes is Kurt McIntire's home-type recognition project, which classified eight US residential styles — Cape Cod, Colonial, Victorian, Tudor, Craftsman, Mediterranean / Spanish, Mid-century Modern, and Edwardian — using ResNet50 fine-tuned on ~100 Bing-scraped images per class. Best result: 76% accuracy. The author noted the small dataset and image-quality variance as the binding constraints; they also dropped Edwardian for poor data and replaced Mediterranean with Spanish for visual consistency. The Stanford CS231n project Classifying U.S. Houses by Architectural Style Using CNNs is the analogous student-project baseline with similar conclusions.

Bottom line on supervised classifiers: A dedicated CNN trained on ~1,000 well-curated US-residential images per class can probably hit 75–85% top-1 accuracy on a 10-12 class US-residential label set, per the converging evidence from McIntire (76%), the CS231n project, and the modern attention-CNN papers. This is a usable baseline — but data collection (>10K labeled US home photos) is real work, and the result is a static model that needs retraining as fashions shift.

A.2 Real-estate datasets and academic datasets

Real-estate listings (Zillow, Realtor.com, Redfin) are a tempting data source — millions of US home photos, often with style metadata, address fields, MLS-derived attributes. None of the three platforms publish public APIs for image bulk download. Their AVMs reportedly use 500+ data points per property (per the Redfin/Zillow comparison) but image-derived style is not a publicly exposed field. Scraping Zillow listings is a known practice but legally fraught and quality- variable (interior shots dominate, exterior shots are inconsistent in framing).

Academic datasets we can use without scraping:

| Dataset | Scale | Classes | Notes | |---------|-------|---------|-------| | Architectural Style (Xu/ECCV) | ~4,800 | 25 (world) | Standard benchmark; uneven classes | | Roboflow ArchiStyles | ~7,500 | 25 | Modern pretrained model included | | Places365 | ~1.8M | 365 scene categories | Has "house" subcategories, not style-specific | | Open Images | 9M | 600 object classes | Useful for object grounding, not style |

For SimplyScapes' purposes — a yard photo with the house partially visible — none of these datasets are directly fit-for-purpose. Architectural Style and ArchiStyles assume a frontal architectural photograph. Real Visual Designer photos are taken from a yard angle, often with the house occupying 20-50% of the frame, sometimes truncated by trees. This is the framing problem (see A.5 below).

A.3 Frontier vision LLM capability — the practical recommendation

Given the data-collection cost of training a custom classifier and the recency of vision LLMs, the practical recommendation is to use a frontier vision LLM with a curated few-shot prompt, not to train a custom CNN. Three reasons:

1. It works at acceptable accuracy out of the box. Few-shot vision LLMs (Gemini 2.5 Pro Vision, GPT-5 Vision, Claude Opus 4 Vision) are strong fine-grained zero-shot classifiers when given a curated set of label exemplars. The MMMU benchmark shows Gemini 2.5 Pro at 79.6% / GPT o3 at 82.9% / Claude Opus 4 at 76.5% on cross-domain visual reasoning. The medical imaging benchmark (Brain MRI sequence classification) shows GPT-4o at 97.7% and Gemini 2.5 Pro at 93.1% on a domain- specific 10-class problem with no domain training. Architectural style is a less-clean signal than MRI but a much smaller class set — probably 12 US-residential classes. A reasonable expected accuracy is 70–85% top-1 with a curated 3-5 exemplar few-shot prompt.

2. The pipeline already runs vision LLMs. Site analysis already calls Gemini Vision for yard type, orientation, climate zone, size, and removable objects. Adding home-style classification is one more field on the same vision call — it does not add a new system dependency. If we already trust Gemini Vision for "is there a driveway in this image", we can trust it for "does this look more Craftsman than Ranch".

3. It evolves with the model. A custom classifier requires retraining every 12–18 months as architectural fashions shift. A prompted vision LLM benefits from every model upgrade with no retraining cost. Gemini 2.0 → 2.5 → 3.0 is a free quality boost.

The proposed label set (12 US-residential classes, matching the existing 12 design systems where the mapping is sensible):

| Home class | Typical features | Recommended design systems | |-----------|------------------|----------------------------| | Craftsman / Bungalow | Exposed rafters, tapered columns, low-pitch roof | Cottage, Naturalistic, Rustic | | Ranch | Single-story, low horizontal lines, attached garage | Modern Farmhouse, Xeriscape, Modern | | Modern / Contemporary | Flat or shed roof, large glazing, clean lines | Modern, Japanese Zen, Xeriscape | | Colonial / Traditional | Symmetric facade, central door, multi-pane windows | Formal Traditional, Cottage | | Tudor | Steep roof, half-timbering, masonry | Formal Traditional, Rustic | | Mediterranean / Spanish | Stucco, terracotta tile, arched openings | Mediterranean, Xeriscape | | Mid-Century Modern | Low-pitch / butterfly roof, post-and-beam, large glass | Modern, Japanese Zen | | Farmhouse / Modern Farmhouse | Gabled roof, board-and-batten, porch | Modern Farmhouse, Cottage, Prairie/Native | | Cape Cod | Steep gabled roof, central chimney, dormer windows | Cottage, Coastal | | Victorian | Asymmetric, ornate trim, turrets, porches | Cottage, Formal Traditional | | Coastal / Beach | Light siding, raised foundation, wraparound deck | Coastal, Tropical | | Other / Unrecognized | Catch-all | (no recommendation; rely on questionnaire) |

The "Recommended design systems" column is what surfaces as a "Recommended for your home" ribbon on the picker — top 2-3 cards.

A.4 Confidence-thresholding patterns

Vision LLMs return per-label confidence (or can be prompted to). The UI needs three thresholds, calibrated to the expected error rate:

| Confidence | UI behavior | |-----------|-------------| | High (>0.80) | Silent application — show "Looks like a Craftsman home" as a passive label on the picker, no confirm step | | Medium (0.50–0.80) | Soft prompt — "Looks like Craftsman. Confirm or change." with a chip menu (one-click correction to top-3 alternates) | | Low (<0.50) | Skip the recommendation — fall through to the no-prediction state (questionnaire / picker without ribbon) |

The "Confidence UI" pattern from Modexa's analysis is the model — show uncertainty as actionable, not embarrassing. Don't display raw percentages; bucket into "looks like" / "could be" / "unsure" affordances.

The threshold values above are starting estimates. Calibration should happen on a labeled validation set of ~200 real Visual Designer photos before launch. Per LlamaIndex and the synchronous-approval pattern, the high-confidence threshold should be set so that fewer than ~10% of "silent" applications turn out to be wrong; otherwise the user's trust erodes. If the model's calibration is poor, default the medium-confidence threshold to 0 (always confirm) until validation data exists.

A.5 The "yard photo with house partially visible" edge case

Real Visual Designer photos are not architectural photographs. They are:

• Taken from the user's yard (not the street)
• Often at an oblique angle to the home
• Frequently with the home occupying 20-50% of the frame, sometimes truncated by trees, fences, or the photo edge
• Sometimes with the home not visible at all (parkstrip, garden bed, side yard)

This breaks both the academic datasets (which assume frontal architectural photographs) and the McIntire-style classifier (which assumed clean front-elevation photos).

Recommended approach for SimplyScapes:

1. Detect home presence first. A pre-filter step asks: "is there a house visible in this image? What fraction of the frame?" If the home is <15% of frame or absent, skip home-style classification and defer to the questionnaire. This is a binary classification the vision LLM handles easily.

2. Few-shot the LLM with oblique-angle exemplars. Build the reference set of 3-5 exemplar photos per class using yard-angle photos, not front-elevation photos. Include the partial-visibility case explicitly in exemplars. This shifts the classifier's distribution to match the deployment distribution.

3. Increase the medium-confidence threshold for partial-visibility cases. If the home is truncated or angled, demote a "high-confidence" classification to "medium-confidence" and surface a confirm step. The vision LLM can be asked to return both confidence and visibility ("home is fully visible / partially visible / truncated") as structured fields.

4. Allow the user to upload a separate front-elevation photo as a reference image (this is the multi-image references feature — see multi-image-design-patterns.md). Power users get more accurate classification by uploading a clean home photo. The first-time user does not have to do this; their yard photo is enough for a soft recommendation.

A.6 Latency and cost — adding to the existing vision call

Because the existing site-analysis pipeline already runs Gemini Vision on the yard photo, adding home-style classification is a marginal-cost extension: extra fields in the same response. Per the Gemini API pricing referenced in report.md § 6.2, this is roughly free. The extra fields:

{
  "home_visible": true,
  "home_fraction_of_frame": 0.32,
  "home_style": "Craftsman",
  "home_style_confidence": 0.78,
  "home_style_alternates": ["Bungalow", "Farmhouse"],
  "existing_landscape_style": "Bare lawn",
  "existing_landscape_confidence": 0.91
}

A single Gemini 2.5 Pro Vision call with this structured output and a few-shot prompt of 3-5 exemplars per class fits well within the existing site-analysis budget.

---

B. Existing Landscape Style Classification

B.1 The maturity gap

Home-style classification has decades of academic prior art. Landscape- style classification has almost none. The closest published work is a single 2026 Nature Scientific Reports paper, "Decoding garden design language via semantic segmentation for social aesthetic interaction," which built a dataset of ~2,000 garden images with pixel-level semantic annotations and aesthetic labels. The paper's contribution is a segmentation approach (DesignSegNet) that produced clearer separation between style clusters (silhouette score 0.68 vs. 0.52 for baseline), but it is fundamentally a research artifact — there is no shipped pretrained classifier, no open-source release, and no validation on real-estate-style photographs.

The vegetation-classification literature exists (UAV multispectral active-learning CNN, urban garden monitoring) but solves a different problem — identifying species or land-cover types from aerial / multispectral imagery. None of it generalizes to "is this a Cottage-leaning yard or a Bare-lawn yard" from a ground- level cell-phone photo.

Conclusion: there is no off-the-shelf pretrained landscape-style classifier we can buy or download. We have to build it ourselves, and the right tool is a curated few-shot vision-LLM prompt, not a CNN.

B.2 Proposed vision LLM approach

Same architecture as A.3, different label set. The vision LLM is asked to return a soft membership distribution across the 12 design systems we ship pre-baked, plus a "what's worth keeping" sub- classification. Two outputs from a single vision call:

Output 1 — Soft membership across the 12 design systems:

{
  "landscape_style_distribution": {
    "Cottage": 0.42,
    "Naturalistic": 0.21,
    "Rustic": 0.15,
    "Bare lawn": 0.10,
    "Modern Farmhouse": 0.08,
    "Other": 0.04
  },
  "primary_label": "Cottage-leaning",
  "confidence_in_primary": 0.42
}

A soft distribution is more useful than a hard label. The 12-system label set is a continuous aesthetic space, not a discrete one — real yards are mixed. Surfacing "Cottage-leaning, with Naturalistic elements" lets the UI show two relevant systems on the picker, not one. Soft membership is also a cleaner input for the prompt- construction layer when the user later picks a system: "user picked Modern Farmhouse, current yard is Cottage-leaning → emphasize foundation-shrub replacement and lawn-shape changes in the prompt".

Output 2 — "What's worth keeping" sub-classification:

{
  "keep_assessment": {
    "mature_trees": { "present": true, "health": "good", "remove_recommendation": "keep" },
    "hardscape": { "present": true, "condition": "fair", "remove_recommendation": "consider replacing if mismatched with chosen design system" },
    "recent_plantings": { "present": false },
    "lawn": { "present": true, "condition": "patchy", "fraction_of_frame": 0.45 },
    "overgrown_or_dead": { "present": true, "items_visible": ["overgrown shrub at right of frame"] }
  }
}

This output feeds Step 1 (Prepare site). The "Prepare site" step already lets the user click objects to remove; this output gives the default keep / remove suggestions that pre-fill that step. A user with mature trees and decent hardscape gets "keep these, remove the overgrown shrubs." A user with bare lawn gets "you're starting fresh — let's go."

The "Prepare site" suggestions track the "existing landscape style" classification: a Bare-lawn yard prompts "start fresh" copy; an Overgrown yard prompts "let's clear what's not working"; a Cottage- leaning yard prompts "let's keep what's already working."

B.3 Validation strategy

Because there is no benchmark, validation has to be internal. Two parallel paths:

1. Style validation. Build a labeled validation set of ~100 real Visual Designer yard photos hand-tagged with primary style. Compare the LLM's primary-label output against the ground truth. Target: primary-label accuracy >70%, primary + secondary accuracy >85%.

2. Keep/remove validation. Build a labeled validation set of ~50 real yard photos with hand-annotated "keep / remove" decisions per visible feature. Compare the LLM's keep_assessment output against the annotations. Target: per-feature keep/remove accuracy >80%.

These thresholds are generous because the downstream UI is forgiving — a wrong keep/remove default is a one-click correction, and a wrong primary-style label is a different design-system ribbon at most. The worst-case failure mode is a user who ignores the recommendation and proceeds normally, which is the same as the no-classification baseline.

B.4 The "what to keep" decision is where this earns its keep

Most of the value of existing-landscape-style classification is not in the picker (where the home-style classification carries the load) — it is in Step 1 of the multi-step flow. Today's brief frames Step 1 as "Prepare the site — remove unwanted items." Without the existing- landscape classification, this step is a generic prompt: "click things to remove." With the classification, it becomes a directed prompt:

"Looks like a Cottage-leaning yard with mature trees and overgrown shrubs. Want to keep the trees and clear the shrubs? You can adjust below."

[✓ Keep mature trees] [✗ Remove overgrown shrubs] [✓ Keep stone path] [— Add to or change]

This is the difference between "the AI doesn't know your yard" (current state) and "the AI noticed you have mature trees and is suggesting we keep them" (proposed state). It is the single biggest UX win of the two classifications, because it converts an open canvas into a tailored checklist with one click of work removed.

B.5 Edge cases

• No yard at all. Concrete-only courtyard, balcony, parkstrip scenarios. The classification should return "Hardscape-only" or "Negligible vegetation" as a primary label with 1.0 confidence and skip the design-system distribution. The UI should detect this upstream of the vision call (yard-type field already covers this).
• Multiple distinct zones in one photo. Front + side yard, lawn + garden bed. The vision LLM can be prompted to return per-zone distributions. Not a v1 priority — v1 should treat the photo as a single zone and let the user crop / re-shoot for a different zone.
• Seasonality. A winter photo of a perennial garden looks like a Bare yard. The classifier should return a "season" field (winter | dormant | growing | peak) and the prompt-construction layer should not over-weight current-state plantings on a winter photo. EXIF date is a fallback; visual cues (snow, dormant grass) are the primary signal.
• Photo quality. Blurry, low-resolution, or off-axis photos degrade classification. The pre-filter from A.5 (home_fraction_of_ frame, photo_quality fields) extends here — return "low quality" as a structured field and skip the soft recommendation when triggered.

---

C. Recommended approach for SimplyScapes (synthesis)

| Task | Approach | Why | |------|----------|-----| | Home architectural style | Few-shot Gemini 2.5 Pro Vision, 12 US-residential classes | No data-collection cost, evolves with model, fits in existing site-analysis call | | Existing landscape style | Few-shot Gemini 2.5 Pro Vision, soft distribution over 12 design systems | No academic baseline exists; soft distribution is more useful than hard label; feeds Step 1 | | Confidence thresholds | High (>0.80) silent, Medium (0.50–0.80) soft confirm, Low (<0.50) skip | Calibration-pending; tighten on validation set of ~200 photos | | Edge cases | Pre-filter (home presence, photo quality, season) returned alongside style | Prevents bad recommendations from misframed inputs | | Validation | Internal labeled validation sets of ~200 home + ~100 landscape + ~50 keep/remove photos | No public benchmark exists; we're the benchmark setter |

The single biggest insight of this delta is that for both classifications, the right tool is a few-shot vision-LLM prompt that runs as part of the existing site-analysis call, not a custom CNN. That decision saves a quarter of engineering work, eliminates a model- ops surface, and makes both classifications upgrade-with-the-model instead of upgrade-by-retraining.

---

D. Open questions for the UX phase

1. How to surface home-style on the picker. A "Recommended for your home" ribbon on the top 2-3 design-system cards is the brief's stated direction. Open: should the recommendation be dismissable? ("Don't recommend by home style" toggle.) Should it persist after a user picks a non-recommended system, or fade?
2. Correction UX for partial misclassifications inside ~300px. The brief calls out home-style and existing-landscape-style as confirmable / correctable. Open: chip menu (top-3 alternates), full dropdown (all 12), or modal sheet with picker grid? Trade-off between density and discoverability.
3. When to confirm vs. silently apply. The threshold values in A.4 are starting estimates. Real numbers come from calibration data we don't yet have. The UX flow should accommodate either path — "confirm" must be an unobtrusive 1-click moment, not a blocking gate.
4. Soft distribution vs. primary label. Existing-landscape-style returns a soft distribution. Does the UI surface that ("70% Cottage, 20% Naturalistic"), or pick the primary and surface the rest as "alternate vibes"? Probably the latter — non-designers don't read distributions.

---

Sources

| # | Type | Reference | URL | |---|------|-----------|-----| | 1 | Paper | Xu et al., "Architectural Style Classification Using Multinomial Latent Logistic Regression" (ECCV 2014) | https://link.springer.com/chapter/10.1007/978-3-319-10590-1_39 | | 2 | Paper | "Architectural style classification based on CNN and channel-spatial attention" (Springer 2022) | https://link.springer.com/article/10.1007/s11760-022-02208-0 | | 3 | Project | McIntire — "Building an image recognition model to classify home types" | https://kurtmcintire.com/image-recognition-homes/ | | 4 | Project | Stanford CS231n — "Classifying U.S. Houses by Architectural Style Using CNNs" | https://cs231n.stanford.edu/reports/2017/pdfs/126.pdf | | 5 | Dataset | Architectural Styles dataset (Kaggle / Xu) | https://www.kaggle.com/datasets/dumitrux/architectural-styles-dataset | | 6 | Dataset | Roboflow ArchiStyles | https://universe.roboflow.com/mipt4/archistyles | | 7 | Paper | "Decoding garden design language via semantic segmentation" (Nature Sci Rep 2026) | https://www.nature.com/articles/s41598-026-46120-w | | 8 | Paper | "Identifying and Monitoring Gardens in Urban Areas Using Aerial and Satellite Imagery" (MDPI 2023) | https://www.mdpi.com/2072-4292/15/16/4053 | | 9 | Paper | "Utilizing active learning and attention-CNN to classify vegetation" (Nature Sci Rep 2024) | https://www.nature.com/articles/s41598-024-82248-3 | | 10 | Benchmark | LM Council — Vision benchmark comparison May 2026 | https://lmcouncil.ai/benchmarks | | 11 | Paper | Brain MRI sequence classification — GPT-4o vs Claude 4 vs Gemini 2.5 Pro | https://pmc.ncbi.nlm.nih.gov/articles/PMC12345967/ | | 12 | Paper | "Zero-Shot Fine-Grained Image Classification Using Large Vision-Language Models" (arXiv 2510.03903) | https://arxiv.org/abs/2510.03903 | | 13 | Paper | "GPT4Vis: What Can GPT-4 Do for Zero-shot Visual Recognition?" | https://arxiv.org/html/2311.15732v2 | | 14 | Article | Modexa — "The Confidence UI Pattern That Users Actually Trust" | https://medium.com/@Modexa/the-confidence-ui-pattern-that-users-actually-trust-ff27e1a8a956 | | 15 | Article | Orkes — "Human-in-the-Loop in Agentic Workflows" | https://orkes.io/blog/human-in-the-loop/ | | 16 | Reference | LlamaIndex — Confidence Threshold | https://www.llamaindex.ai/glossary/what-is-confidence-threshold | | 17 | Docs | Gemini API — Few-Shot Examples | https://cloud.google.com/vertex-ai/generative-ai/docs/learn/prompts/few-shot-examples | | 18 | Docs | Roboflow — Prompting Tips for LLMs with Vision | https://blog.roboflow.com/prompting-tips-for-large-language-models-with-vision/ |

---

Version History

| Date & Time (MT) | Author | Summary | |-------------------------|-------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | 2026-05-01 11:00 PM MT | Research Analyst | Initial draft. Covers Deltas 1 & 2: home architectural style and existing landscape style classification. Recommends few-shot vision LLM over custom CNN; defines confidence thresholds; addresses yard-photo edge case; identifies "what to keep" Step 1 win as biggest UX impact. |

Multi-Image, Reference-Image, and Cross-Area Design Profile Patterns

Date: 2026-05-01 | Status: draft Parent report: research/report.md Part VI Idea: generative-ai-chat-interface

This document covers Delta 3 of the 2026-05-01 research round — how adjacent tools handle (a) multi-photo input, (b) reference images for style or structure, and (c) cross-area / cross-room design profiles. The brief introduces a Property Design Profile and three reference- image subtypes (alternate base view, property reference, aspirational). This research grounds the panel-internal UX design for those features in proven patterns from interior-design tools and frontier image- generation interfaces.

The deepest analogues are interior design — Modsy, Havenly, Decorilla, Houzz, Pinterest. They solve the exact same problem (photo of a real space + style preference + reference inspiration → designed render) at a more mature point in the curve. Frontier image-gen tools (Midjourney, Adobe Firefly, ReRoom, Spacely) handle the technical reference-image patterns; interior-design tools handle the cross-room profile pattern.

---

A. Modsy — the closest dead analogue

Modsy was the textbook example until it shut down virtual design services in 2022. Its mechanics are still the canonical interior-design quiz / multi-photo / cross-room flow. Source: Modsy review (Decorilla) and the still-live Modsy style quiz.

The quiz is image-pick, not text. Per the live quiz page: "All you do is look at a few images and tell us which you like and which you don't." Users browse rooms and react. The algorithm reads patterns in their picks. The quiz takes 2-6 minutes.

Output is a blended profile, not a single label. "Modsy rejects the one-style-fits-all approach, instead providing a unique, blended result that's incredibly personal." A typical output recognizes "Mid-Century Modern furniture preference alongside industrial materials" — a soft distribution. Strong precedent for our recommendation in home-style-classification.md § B.2 that landscape-style classification should return a soft distribution rather than a hard label.

The room-photo upload happens after the quiz, not during. Once a user starts a project, they upload room photos and measurements. Modsy built a 3D model of the user's actual room from those photos. Style is inferred from quiz preferences, not detected from the room photo.

Cross-room design profile: Implicit in the user account, not a formalized entity. Each new room project inherits the user's quiz profile by default. Per-room style override happened in conversation with the assigned designer, not as a UI affordance.

What transfers to SimplyScapes:

• Image-pick quiz over text. Brief Q3 calls out the questionnaire; the existing sample-questionnaire.md already uses image-picks for the home-vibe question. Extend this pattern.
• Soft profile, not hard label. Quiz output is a distribution across systems, ranked. The picker shows top 3-4 systems with ribbons rather than auto-locking one.
• Profile is a per-user / per-property entity, not a per-room one. Modsy proved homeowners think property-first, not room-first.

What doesn't transfer:

• Modsy's per-room style override happened in a designer conversation (out-of-band). SimplyScapes needs a UI affordance for per-area override (e.g., backyard = Tropical, property = Cottage). This is Q7 in the brief.

---

B. Havenly — the live equivalent

Havenly is the still-running Modsy equivalent. Same model, lower-touch: human designer + 3D render, $159 starting price (per Decorilla's Modsy alternatives roundup and Havenly review).

The Havenly style quiz (live link) is fast and visually driven. Users select rooms they like and receive a personalized style profile. Specific question count and mechanics are not publicly documented, but the live quiz follows Modsy's image-pick pattern.

Output is a single-style label ("Bohemian", "Glam", "Midcentury Modern") — different from Modsy's blended approach. The label is used to match the user with a designer whose specialty fits.

Cross-room behavior: the same. Profile is per-user, applied to every new project. Per-room override happens in designer conversation.

What transfers:

• Image-pick quiz at scale (this is the dominant interior-design quiz pattern).
• Profile-applied-to-projects pattern. New projects inherit; explicit override per project.

What doesn't transfer:

• Single-label output is a regression from Modsy's distribution approach. SimplyScapes should follow Modsy here, not Havenly.
• Designer-mediated override is not a self-serve UI pattern. We need a self-serve override.

---

C. Decorilla — two-designer comparison pattern

Decorilla starts the same way (style quiz → designer match) but adds a distinct mechanic: two designers each produce a concept, and the user picks. Per the Decorilla vs Modsy comparison and Decorilla vs Havenly.

The two-concept pattern is functionally equivalent to branching the generation tree — exactly what report.md § V.E already proposes for the version-history-as-concept-comparison surface. Decorilla validates the demand: "show me two takes" is a recurring homeowner request, not an edge case.

What transfers:

• Demand for parallel concepts is real. The report.md § V.E recommendation to elevate version history into concept comparison is reinforced by Decorilla's market position.

What doesn't transfer:

• Decorilla's two concepts come from two human designers with different aesthetics. Our two concepts come from two design-system applications — we need to be explicit that a SimplyScapes "concept" is a system + step combination, not a designer's personality.

---

D. Houzz — multi-image as a first-class affordance

Houzz is the most multi-image-native of the interior-design tools. Source: Houzz Ideabooks tutorial and Houzz quiz.

The Style Quiz is image-rating: users look at room photos and rate them, and the site infers style(s) from the answers. Output is a single or multi-label style description.

The Ideabook is the killer feature. Users save photos to personal ideabooks. One project can have multiple ideabooks — one for layout, one for colors, one for furniture. The user organizes inspiration by intent, not by source. This is a strong precedent for reference image tagging in SimplyScapes — the brief's Q8 asks whether references should be weighted by intent ("use this for color palette" vs. "use this for plant placement"). Houzz proves users will self-organize references by intent if given the affordance.

Multi-room organization: Users create separate ideabooks per room or per intent. There is no formal "design profile" entity in Houzz — ideabooks are the substitute. A user's whole-house aesthetic emerges from the union of their ideabooks rather than from a top-down profile.

What transfers:

• Reference-by-intent tagging is a real pattern. When users organize references, they segment by intent (color, layout, material). The SimplyScapes References section can mirror this with optional tags. Q8 of the brief is essentially asking "should we be Houzz here?" Answer: yes, but as an optional tag, not a required one. Default behavior is "general inspiration"; tags are additive.
• Multiple references per project are normal, not unusual. Users expect to upload 5-20 inspiration images per project. Single- reference UI is the regression.

What doesn't transfer:

• Houzz has no top-down profile. Their model is bottom-up emergence from ideabooks. SimplyScapes wants a top-down Property Design Profile because the prompt-construction layer needs a coherent design-system anchor. The Ideabook pattern is good for the References section; it's not good for the Profile section.

---

E. Pinterest — visual search as the entry point

Pinterest Lens is the most aggressive multi-image input in the ecosystem. Source: Building Pinterest Lens engineering blog and Pinterest's Visual Lens analysis.

Lens does object recognition, scene understanding, and visual similarity matching simultaneously, returning embeddings that drive visual search. The user's input is one photo; the output is N matching inspiration pins. Embeddings are rebuilt nightly.

Boards and sections are the multi-image organization primitive. Per Pinterest's board-creation docs, boards have sections — a user can have a "Backyard 2026" board with sections for "Lighting", "Plantings", "Hardscape". Each section is itself a multi-image collection. Multi-image at multiple levels of hierarchy.

The Taste Graph personalizes. Pinterest builds a "Taste Graph" mapping user preferences from past activity. Search results are personalized — two users searching "living room ideas" see different results based on their saved-pin history.

What transfers:

• Hierarchical multi-image organization (board → section). This is a strong pattern for the SimplyScapes Reference Library. Property- level references vs. design-level references vs. generation-level references map cleanly to a 3-level hierarchy. Don't reinvent hierarchy here — it works.
• Embedding-based similarity for "find me more like this". A v2 feature: when a user saves an inspiration image, surface "similar references already in your library" — emergent across the user's property collection.

What doesn't transfer:

• Pinterest's discovery mechanics (embedding nightly, taste graph scale) are out of scope for v1. The Reference Library can be a manual upload + per-image tag affordance.
• Pinterest has no "this profile drives generation" concept. References are pure input to human discovery, not to AI generation. Our references ARE inputs to generation.

---

F. Frontier image-gen tools — the technical reference-image patterns

Where the interior-design tools handle profiles, frontier image-gen tools handle the per-generation reference-image mechanics. Three patterns dominate.

F.1 Adobe Firefly — typed references (Style + Structure)

Per Firefly style references docs and Firefly structure references:

Firefly distinguishes two reference types as first-class UI affordances:

• Style Reference — "matches the visual style of a reference image (the colors, lighting, textures, mood and theme)." Upload an image, Firefly extracts style and applies it to the generation.
• Structure Reference — "match the composition / layout of a reference image." Upload an image, Firefly extracts spatial structure (the shape and arrangement of objects) and applies it.

Both can be combined in one generation. The UI surfaces them as distinct upload slots in the panel, side-by-side, each with its own upload affordance and preview thumbnail.

Curated style library is also offered: pre-baked style references across categories (3D, Landscapes, Lighting, Texture). User picks one from a grid; same effect as uploading.

What transfers:

• Typed reference slots. SimplyScapes has three reference subtypes (alternate base view, property reference, aspirational). Each deserves its own clearly-labeled slot, not a generic "drop image here" zone. Density: each slot is collapsed by default and expands to show thumbnails of attached images.
• Library + upload, not just upload. SimplyScapes already has a property-level library (existing visual designs from this property, alternate base views). Surfacing those in the references picker is a pure win — most users will reach for "use the front-yard design as a reference for the backyard" before they reach for "upload a new inspiration photo."

What doesn't transfer:

• Firefly's Style and Structure are too technical for the SimplyScapes audience (mostly homeowners, not designers). Use plainer labels: "What's already there" (alternate base view), "Other parts of this property" (property reference), "What I like" (aspirational). Match the existing brief language.

F.2 Midjourney — multiple references with implicit roles

Per Midjourney Style Reference docs and Midjourney Omni-Reference:

Midjourney's pattern is flat parameter list, multiple URLs allowed. Style Reference syntax: --sref URL1 URL2 URL3 — multiple references of the same type, separated by spaces. Character Reference (--cref) is similar but for characters.

Omni-Reference (introduced in v7) is a unified reference type that "blends style and character traits automatically." This is interesting: Midjourney concluded that splitting Style and Character into separate parameters created user confusion, and merged them. The trend is toward fewer typed slots, more implicit role inference.

Style Codes (--sref CODE) are pre-baked numerical style fragments. Rather than uploading an image, the user references a code from Midjourney's internal style library. Same effect, lower latency, no upload step.

What transfers:

• Multiple references per generation, no hard cap. Don't artificially limit references. Cost / latency is the natural backpressure (see F.4 below).
• Pre-baked style fragments are a strong primitive. Each of the 12 design systems already has a starter prompt fragment (per ux/design-systems.md § C.3). These are SimplyScapes' equivalent of --sref codes. The system pill plays the role Midjourney's --sref plays.

What doesn't transfer:

• Discord-style command syntax is wrong for a homeowner panel. Hide all parameter syntax behind UI affordances.

F.3 ReRoom AI — multi-angle as a reference subtype

Per ReRoom AI overview and tool list:

ReRoom AI ships Multi-Angle Rendering as a first-class tool — "upload up to six angles for consistent outputs across views." This is exactly the SimplyScapes "alternate base view" subtype. Six angles is a useful upper bound. The other ReRoom tools listed:

• Style Reference (style transfer from one image to another)
• Sketch → Image
• Render Enhance
• Image Upscale 4×
• Image Editor (selective replace/remove)

The fact that ReRoom carved out Multi-Angle as a distinct tool — rather than burying it inside Style Reference — is a signal that homeowners think of "this is the same room from another angle" differently from "this is what I want it to look like." Strong precedent for the brief's three-subtype reference model.

What transfers:

• Multi-angle / alternate-base-view is a distinct first-class pattern. Don't merge it with aspirational references in the UI.
• Six-angle upper bound is a sensible default. Match this in v1.
• Generation-level reference selection. ReRoom lets the user pick which angles are active for which generation. This maps to the brief's "Active References per generation" entity. Implementation: the references panel shows all property-level references; the user toggles which are active per generation.

What doesn't transfer:

• ReRoom is interior-only. The yard angle problem (oblique, partial) doesn't apply to interior shots — they're always frontal-rectangular. SimplyScapes should validate the multi-angle UX with real yard photos that have the home truncating one angle.

F.4 Spacely AI — preset/prompt mode and style transfer

Per Spacely AI website and the Q1 2026 design-system update:

Spacely's UI offers preset style picks (curated grid) AND custom style references (user upload) as parallel paths. The report.md § V already cites Spacely as the model for the "Preset/Prompt mode toggle."

The new (Q1 2026) feature is "AI Design System" — a property-level style collection that propagates across rooms. This is exactly the Property Design Profile pattern from the brief, validated by a live competitor in interior design. Direct precedent.

What transfers:

• Property-level style collection that propagates is shipping in market. Validates the brief's Property Design Profile entity.
• Style transfer from a reference image is mature tech. Implement it in v1.
• Preset + Prompt mode in the same UI is the current best practice. Already in the brief.

What doesn't transfer:

• Spacely's "AI Design System" is interior-only and lighter on architectural rigor than landscaping needs. Our profile carries the questionnaire answers, water/maintenance preferences, climate context — substantially richer payload than Spacely's room-style profile.

---

G. Cross-cutting patterns (synthesis)

Across all eight tools (Modsy, Havenly, Decorilla, Houzz, Pinterest, Firefly, Midjourney, ReRoom, Spacely), four patterns recur:

G.1 Multi-image is normal, not exceptional

Every mature tool supports multiple references per project. The hard question is how to surface them in a panel-constrained UI, not whether to have them at all. Recommended UI for SimplyScapes' ~300px panel:

• References section is collapsible, defaults to collapsed when no references are attached.
• Three subtype-labeled slots (alternate views, property, aspirational), each showing a horizontal scroll of small thumbnails (40-60px square) with a "+" tile to add.
• Upper bound of ~6 references per subtype, ~10 references total per generation. Soft cap (warn) at the limit, not a hard block.
• Per-generation toggle for which references are active. UI: each thumbnail has an on/off state; the user can see which ones the next generation will use.

G.2 Profile is per-property / per-user, not per-area

Every tool that has a profile (Modsy quiz, Havenly quiz, Spacely AI Design System) attaches the profile to the user or the property — never to the individual room or area. The brief's Property Design Profile follows the dominant pattern. The override mechanism (per-area override of profile attributes) is the open UX question (Q7) — see G.3 below.

G.3 Override patterns vary; inheritance affordances are rare

This is the area with the least mature prior art. Modsy/Havenly override per-room through human designer conversation — not a UI pattern. Spacely's AI Design System propagation does not yet expose a clear override UI. Houzz's Ideabook system avoids the problem by having no profile at all.

The closest UI prior art is outside design tools:

• Figma component overrides. A component instance shows "inherits from main component" with explicit override indicators on each property. Users can break the link to "promote" an instance to a standalone component, or "push" an override back to the main component. This is the cleanest existing pattern and is what the brief's Q7 hints at.
• CSS cascade UIs (browser DevTools "Computed" panel, design-tool variable overrides). Show the inherited value, then the override, with a clear way to revert.

For SimplyScapes panel-constrained UI:

• Inheritance indicator at the top of the panel: a small breadcrumb showing "Backyard ← Property Profile (Cottage)" when inherited, or "Backyard (Tropical) — overrides Property (Cottage)" when overridden.
• Override affordance on the system pill: tap pill → menu with "Change for backyard only" vs. "Change for whole property". Two paths from one click.
• "Propagate to property" reverse path: when the user changes the backyard's system, offer "make this the property default?" as a one-tap follow-up. Don't auto-propagate.

This is a thin UI atop a clean data model (per-attribute diffs), not an elaborate inheritance manager. The data model carries the complexity; the UI is just one extra tap.

G.4 Tagging references by intent is a Houzz-only pattern

Only Houzz lets users tag references by intent (color, layout, material). This is an optional power-user feature, not a default. For SimplyScapes:

• v1: untagged references. All references contribute generic "consider this aesthetic" context to the prompt-construction layer.
• v2: optional tags. A small tag chip per reference: "color palette", "plant placement", "hardscape style", "lighting." User applies one or more tags. Prompt-construction layer uses tags to weight per-aspect influence. Adoption is likely <20%, but the power-user value is high.

---

H. Recommended approach for SimplyScapes (synthesis)

| Question | Recommendation | Source | |----------|----------------|--------| | How is multi-image surfaced in the panel? | Collapsible References section; three subtype-labeled slots; horizontal scroll of small thumbnails per slot | Firefly (typed slots), ReRoom (multi-angle as first-class) | | Profile entity scope? | Property-scoped, not area-scoped (already in brief) | Modsy / Havenly / Spacely converge here | | Profile output type? | Soft distribution over design systems, not hard label | Modsy precedent | | Override UI inside ~300px? | Inheritance breadcrumb + system-pill menu with two paths ("change for area only" / "change for whole property") + "propagate up" follow-up | Figma component overrides + CSS cascade | | Per-generation reference selection? | Per-thumbnail on/off toggle; soft cap at ~10 active per generation | ReRoom multi-angle pattern | | Reference tagging? | v1 untagged; v2 optional tags ("color palette", "plant placement", etc.) | Houzz Ideabook intent organization | | Style extraction from reference? | Vision LLM call returns extracted color palette + structural notes alongside the image; the prompt-construction layer uses extracted summary, not raw bytes | Firefly Style Reference + ReRoom Style Reference | | References cost / latency? | Each active reference adds ~25% to the generation latency and ~$0.005 to cost (rough estimate, validate); soft warn at 6 active references | Industry-standard image-gen multi-input pricing |

---

I. Open questions for the UX phase

1. References-section default state. Collapsed when empty, expanded when references attached, or always collapsed with a count? Probably the first, but needs wireframe iteration.
2. Reference upload affordance. Drag-and-drop into the panel, click-to-upload button, or both? Mobile pattern?
3. Per-generation reference selection vs. always-on. Is the "active references" toggle a per-generation choice the user makes each time, or is it sticky across generations? Sticky is easier; per-generation is more flexible. Probably sticky with explicit "use different references" for the current generation.
4. Promoting a generation reference to a property reference. When the user attaches an inspiration image to a generation, should we offer "save this to property references" as a one-tap follow-up? Probably yes — converts a one-off to a persistent reference.
5. Override affordance density inside ~300px. The system pill already lives in the panel. Adding inheritance breadcrumb + override menu is more density. Test whether this can fit cleanly without crowding.

---

Sources

| # | Type | Reference | URL | |---|------|-----------|-----| | 1 | Product | Modsy Style Quiz | https://www.modsy.com/design-style-quiz/ | | 2 | Review | Modsy Review (Decorilla) | https://www.decorilla.com/online-decorating/modsy-review/ | | 3 | Industry | Modsy alternatives 2026 (Remodel AI) | https://www.remodelai.io/blog/modsy-alternatives | | 4 | Product | Havenly Interior Design Style Quiz | https://havenly.com/interior-design-style-quiz | | 5 | Review | Havenly Review (Decorilla) | https://www.decorilla.com/online-decorating/havenly-review/ | | 6 | Review | Decorilla vs Havenly | https://www.decorilla.com/online-decorating/decorilla-vs-havenly-which-online-interior-design-service-is-right-for-you/ | | 7 | Review | Decorilla vs Modsy | https://www.decorilla.com/online-decorating/decorilla-vs-modsy/ | | 8 | Product | Houzz Quiz: Find Your Design Styles | https://www.houzz.com/ideabooks/houzz-quiz | | 9 | Tutorial | Houzz Ideabooks Tutorial (Jillian Lare) | https://jillianlare.com/houzz-ideabooks-tutorial/ | | 10 | Article | "Find Your Design Style" using Houzz Ideabooks | https://www.houzz.com/magazine/8-steps-to-finding-your-design-style-using-ideabooks-stsetivw-vs~178358057 | | 11 | Engineering | Pinterest Engineering — Building Pinterest Lens | https://medium.com/pinterest-engineering/building-pinterest-lens-a-real-world-visual-discovery-system-59812d8cbfbc | | 12 | Article | "Pinterest's Visual Lens: How computer vision explores your taste" | https://towardsdatascience.com/pinterests-visual-lens-how-computer-vision-explores-your-taste-5470f87502ad/ | | 13 | Docs | Pinterest Create — Boards | https://create.pinterest.com/product-features/how-to-create-boards/ | | 14 | Docs | Adobe Firefly — Reference images for styling | https://helpx.adobe.com/firefly/web/generate-images-with-text-to-image/customize-generated-images/reference-images-for-styling.html | | 15 | Docs | Adobe Firefly — Structure Reference | https://www.adobe.com/learn/firefly/web/generative-ai-structure-reference-image | | 16 | Docs | Adobe Firefly — Match image composition to reference image | https://helpx.adobe.com/firefly/web/generate-images-with-text-to-image/customize-generated-images/match-image-composition-to-reference-image.html | | 17 | Docs | Midjourney — Style Reference | https://docs.midjourney.com/hc/en-us/articles/32180011136653-Style-Reference | | 18 | Docs | Midjourney — Omni-Reference | https://docs.midjourney.com/hc/en-us/articles/36285124473997-Omni-Reference | | 19 | Guide | Midjourney Omni-Reference Guide (ImaginePro) | https://www.imaginepro.ai/blog/2025/7/midjourney-omni-reference-guide | | 20 | Product | ReRoom AI | https://reroom.ai/ | | 21 | Article | ReRoom AI overview (Creati.ai) | https://creati.ai/ai-tools/reroom/ | | 22 | Product | Spacely AI | https://www.spacely.ai/ | | 23 | Product Update | Spacely AI Q1 2026 — AI Design System | https://resources.spacely.ai/whats-new-in-spacely-ai-your-complete-ai-design-system-has-arrived-q1-2026-edition/ | | 24 | Product | Higharc Studio — Generative home plan design | https://www.higharc.com/product/studio | | 25 | Product | Maket — AI floor plan creation | https://www.maket.ai/ | | 26 | Article | Maket — AI Floor Plan Generator Guide 2026 | https://www.maket.ai/blog/ai-floor-plan-generator-guide-2026 | | 27 | Product | Yardzen — How It Works | https://yardzen.com/how-it-works |

---

Version History

| Date & Time (MT) | Author | Summary | |-------------------------|-------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | 2026-05-01 11:30 PM MT | Research Analyst | Initial draft. Covers Delta 3: multi-image, reference-image, and cross-area design profile patterns. Eight-tool comparative scan (Modsy, Havenly, Decorilla, Houzz, Pinterest, Firefly, Midjourney, ReRoom, Spacely). Recommends collapsible References section with three subtype slots, property-scoped profile, soft-distribution output, Figma-style override UI. Open questions surface for the UX phase. |